How to Become a Builder PM (n8n, Claude Code, OpenClaw)

频道: Aakash Gupta
视频: https://www.youtube.com/watch?v=PL7908aNeSE
原文语言: en
统计: 共 67 轮 · Mahesh 49 · Aakash 8

[0:00] Mahesh

I would love to send this message to all PMs that this is our time to shine. There is a lot of misconception there that, you know, if you start using Claude code or if you configure open claw, you become a builder PM. Mahesh Yadav, who's been a PM everywhere. Microsoft, Amazon, Meta, last not least, Google.

我特别想把这句话送给所有 PM：这是我们大显身手的时候了。外面有个很大的误解，就是觉得你只要开始用 Claude Code、配好 OpenClaw，你就成了 builder PM。Mahesh Yadav，哪儿都干过 PM——Microsoft、Amazon、Meta，还有最后但同样重要的 Google。

[0:19] Aakash

This is the time to build your own world, and I believe in that future, and that's why I left, and I have zero regrets. If you know your skills, you can build it in Claude code and delegate that work to agents. I don't know its limitations. It does anything that I want it to do. This is an open challenge I have given to anybody that if you can do a job, just tell me the job, and this agent will do it better than you, 365 days, 24 hours. Everything which used to take you almost two to three months for us to write the PRD, to get to mocks, from mocks to a real working prototype, from there to getting customers and seeing the signals, all that is getting squeezed with this Claude code, and you become the builder PM that the world needs. The ability to sandbox these agents in a controlled way, that's an unsolved problem, and that I think is what I am excited about. Google isn't going to allow you to just give your company access to an open claw. Yeah, no, I 100% agree. I think the idea is Before we go any further, do me a favor and check that you are subscribed on YouTube and following on Apple and Spotify podcasts. And if you want to get access to amazing AI tools, check out my bundle. Where if you become an annual subscriber to my newsletter, you get a full year free of the paid plans of Mobb arise, relay app, dovetail, linear, magic patterns, deep sky, reforge, build, descript, and speechify. So, be sure to check that out at bundle.akashg.com, and now into today's episode. PMs are now are being asked to push PRs. PMs are are being asked to code. This is the rise of the builder PM. But, what is a builder PM? Today, I have Mahesh Yadav, and today he's going to help you understand how do I become a builder PM? How do I use N8N? How do I use Claude

现在是你打造自己天地的时候了，我相信那个未来，所以我才离开，而且我一点都不后悔。只要你清楚自己的技能在哪，你就能在 Claude Code 里把它造出来，然后把活儿交给 agent 去干。我看不到它的天花板，我想让它干什么它就能干什么。我公开放过话：只要你能干的活儿，把那活儿告诉我，这个 agent 都能比你干得更好，一年 365 天、一天 24 小时不停。以前我们写 PRD、做出 mock、再从 mock 做出能跑的真实原型，再拿去找客户、看市场反馈，这一整套差不多要花两三个月，现在用 Claude Code 全被压缩了，你也就成了这个世界需要的 builder PM。怎么用一种受控的方式给这些 agent 套个沙箱，这还是个没解决的难题，而这恰恰是我兴奋的点。Google 是不会让你随便把公司权限交给一个 OpenClaw 的。对，没错，我百分百同意，我觉得这个思路……在我们往下聊之前，帮我个忙，确认一下你已经在 YouTube 订阅、在 Apple 和 Spotify 播客上关注了我们。如果你想用上那些超棒的 AI 工具，去看看我的套餐包：只要你成为我 newsletter 的年度订阅会员，就能免费用一整年这些工具的付费版——Mobb、arise、relay app、dovetail、linear、magic patterns、deep sky、reforge、build、descript 还有 speechify。一定去 bundle.akashg.com 看看。现在进入今天这期节目。现在 PM 被要求去提 PR，PM 被要求去写代码，这就是 builder PM 的崛起。但什么是 builder PM？今天我请来了 Mahesh Yadav，他会帮你搞懂：我怎么才能成为 builder PM？怎么用 n8n？怎么用 Claude

[2:10] Aakash

code? How do I use open claw in order to become a more effective and efficient PM, even if I'm not building AI feature. Mahesh, everybody loved our last episode. Thanks for coming back. Oh, thank you for having me, and I think this is the time of urgency. So, I would love to use your platform to send this message to all PMs that uh this is our time, and we should be ready when this time arrived. I was always preparing for this time, and now the time is right for all the PMs to shine. It's just a little bit that we need to go learn, and if we learn what we need to learn, this is our moment. So, what is a builder PM, and how does a PM become one?

Code？怎么用 OpenClaw，好让自己成为一个更高效、更顶用的 PM——哪怕我做的根本不是 AI 功能。Mahesh，大家都特别喜欢我们上一期，谢谢你再次过来。哦，谢谢你请我来。我觉得现在是个紧迫的时刻。所以我特别想借你这个平台，把这句话传达给所有 PM：这是属于我们的时刻，时机一到我们就得准备好。我一直在为这一刻做准备，现在正是所有 PM 大放异彩的时候。只是我们还得去学一点东西，而只要我们学会了该学的，这就是我们的高光时刻。那么，什么是 builder PM，一个 PM 又该怎么成为 builder PM？

[2:47] Mahesh

Ah, that's a very good question, right? I was always right, means I had an engineering background. I was not a traditional PM who came from a B school and then went to McKinsey and then became a PM. I had a very gradual move to PM. I was an engineer, and I was always building, and then I became a PM because I was always building what is customer wanted or working backward from customers, rather than just building for the sake of building, which is very popular at Microsoft, if you don't know. So, for me, a builder PM is like as PMs, we are always building. Our job is to build the right thing. And now, earlier, if you had the tools, you would have built the whole product, but it was very hard to build anything, or you need at least three or six months of rigorous coding, testing, deployment, all that was needed to go build things, but in the new age, even like people who are engineers all alongs are saying that they're not writing code anymore. They are talking to customers, and the Claude code does coding for them. In that age, the skill that becomes important is like what to build and what does customer want, and that you have. And if you use on the right-hand side, the tools to do the right prototyping and then build at least the first version of your pro- product, then you become the builder PM that the world needs, and I think all of us need to grow into that, whether engineers, designers, or PMs, because without that, we will not be able to diffuse the benefits of AI into economy. So, for me, a builder PM is somebody who is who has taken the responsibility to diffuse the benefits of this awesome AI we have today into econ- into the economy so that we can all derive the benefits that is that large companies and research labs are putting so much money into to build.

啊，这个问题问得特别好。我可以说一直都站对了方向，意思是我有工程背景。我不是那种从商学院出来、先去麦肯锡、再转 PM 的传统 PM。我转 PM 是一步步慢慢来的。我本来是工程师，一直在动手造东西，后来成了 PM——因为我一直在造客户真正想要的东西，也就是从客户需求倒推着做，而不是为了造而造，后者在 Microsoft 可是相当流行，怕你不知道。所以对我来说，builder PM 就是——作为 PM，我们一直都在造东西，我们的活儿就是把对的东西造出来。以前哪怕你有工具，你也得自己把整个产品做出来，可那时候造任何东西都特别难，至少得三到六个月的严苛编码、测试、部署，全套都得来一遍才行；但到了新时代，连那些一直在写代码的工程师都说自己已经不再写代码了，他们在跟客户聊，Claude Code 替他们写代码。在这个时代，变得重要的技能是——造什么、客户到底想要什么，而这个本事你本来就有。如果你再在右手边用上对的工具去做对的原型，把产品的第一个版本至少先造出来，你就成了这个世界需要的 builder PM。我觉得我们所有人都得朝这个方向成长，不管是工程师、设计师还是 PM，因为没有这个，我们就没法把 AI 的红利扩散到整个经济里去。所以对我来说，builder PM 就是那个主动扛起责任、把我们今天拥有的这个了不起的 AI 的红利扩散到经济里去的人，好让我们都能享受到那些大公司、研究实验室砸了那么多钱才造出来的成果。

[4:46] Mahesh

So, for me, nutshell, build PM is somebody who can take talk to customers, figure out what needs to be built, and build the first version, and get to 10 customers without talking to an any developer at all.

所以对我来说，一句话概括，builder PM 就是这么个人：能跟客户聊，搞清楚该造什么，把第一个版本造出来，然后做到 10 个客户，全程根本不用跟任何一个开发者说一句话。

[4:59] Aakash

Amazing. Can you show us in action what are the skills and concepts we need to understand in order to get there?

太棒了。你能不能现场给我们演示一下，为了走到那一步，我们需要搞懂哪些技能和概念？

[5:04] Mahesh

Yeah, I think there is a lot of misconception there that, you know, if you start using Claude code or if you configure open claw, you become a builder PM. Uh I think it's the the hype is right, because this is the first time people see that they can manage their calendar, they can delegate to work to other party, or they can just say things, and it just happens. But, AI, my my I've been in AI for last 10 years and build things all along. I think just knowing the layers or understanding how these things work is the first step or of building these things. And for me, I start my journey with an obviously you talked about a lot of tools. So, I will start my journey in first understanding these concepts. So, I will start in earlier day, I will start with something like tensor flow or pytorch and understand what a model is or train a model, and then inference a model. In the new world, in the agentic AI, I will start with something like N8N, and then go and say, "Hey, okay, what is an agent? How it interacts with model? What is a model? What are the limitation of it? What is memory? What is tools? And maybe I can just show you if it's cool. Maybe I just share my screen and show you because it's not very hard to learn these things in N8N. N8N has done a good job, to be honest. Uh maybe it's an obsolete tool in building workflows now with Claude code, but I think it's still an amazing tool to learn. So, let me just share quickly with you like what it takes to build or what it takes what are the components involved in building these agents. So, just if uh a revision revision from last time. If you look at last time, right? We talked about this idea that as we grow as humans, kid is born, they first need to have the knowledge of the world. So, this is how the world works.

好，我觉得外面有个很大的误解，就是觉得你只要开始用 Claude Code、配好 OpenClaw，你就成了 builder PM。呃，我觉得这股热度本身没毛病，因为这是人们第一次发现自己能管理日程、能把活儿派给别人、或者就只是动动嘴说一句，事情就办成了。但 AI——我在 AI 领域干了过去 10 年，一直在动手造东西。我觉得，光是知道这些层、搞懂这些东西是怎么运作的，这才是造这些东西的第一步。对我来说，我开启这趟旅程……你显然提到了很多工具，所以我会先从搞懂这些概念入手。早些年我会从 TensorFlow 或者 PyTorch 这类东西开始，搞懂什么是模型、怎么训练一个模型、怎么对模型做推理。在新世界、在 agentic AI 里，我会从 n8n 这类东西开始，然后去问：哎，好，什么是 agent？它怎么跟模型交互？什么是模型？它有哪些局限？什么是 memory？什么是 tools？要是方便的话我可以直接演示给你看，我把屏幕分享出来给你看，因为在 n8n 里学这些东西其实不难。说实话 n8n 做得挺好。呃，现在有了 Claude Code，拿它搭 workflow 也许已经算过时工具了，但我觉得拿它来学习还是个超棒的工具。那我快速跟你分享一下，搭一个这样的东西需要什么、搭这些 agent 都牵涉到哪些组件。先复习一下上次的内容。回想上次，对吧？我们聊过这么个想法：我们人类成长的过程——一个孩子出生，他首先得有对这个世界的认知，也就是这个世界是怎么运转的。

[7:12] Mahesh

Great. Okay, I understand that, and I can build my intelligence on that, and I can also update it. And second, I need you to understand what is the current state of the world. So, if I want to do anything, I want to know we teach our kids all the time. When they are born, we will say to them that, "Hey, this is hot. This is cold." All those signals, which is current state of the world. Yeah, we know that this is a thing that gives us power or this is where we cook, but this is the current state of the world. So, this is your signals or memory tells you what is the current state of the world. And then, if you did a good job, you can ask your kids to get a glass of water for you, and that's the tools where they can use a tools like a gla- glass, and then open a tap, and then move, hold it, and get it back to you. So, that's the tools piece, and then we learn the guardrails, the laws. What is possible in this country? If you're moving on a road, then you need to look left and right. It's your responsibility, not the responsibility of the driver, especially on a high- especially on a busy road. So, we tell our guardrails and laws. Yeah. If we do all this, and if we just replace humans here with a model, so this model is analogous to what you get from OpenAI or Anthropic or Google, the model is just the intelligence layer. It just trained to predict the next word, and now have some reasoning, but you need all this harness, harness, or people call it scaffolding, to actually build something that can solve problems or build impact for you. So, this harness is called agents, and then we use these frameworks like N8N to build these agents, right? And then, if you look at every agent that we have built so far or what people have built has one of these four things in it, or

很好。好，我懂了这个，我能在这之上建立起我的智能，我也能更新它。第二，我得让你理解这个世界当下的状态。所以，我要想做任何事，我都得知道——我们一直在教自己的孩子，他们一出生，我们就会跟他们说：哎，这个烫，这个凉。所有这些信号，就是这个世界当下的状态。对，我们知道这是个能给我们能量的东西，或者这是我们做饭的地方，但这就是世界当下的状态。所以这就是你的信号，或者说 memory，它告诉你世界当下是什么状态。然后，如果你教得好，你就可以让孩子去给你倒杯水，这就是 tools——他们会用到杯子这样的工具，然后拧开水龙头，再端着、扶稳、送回到你手上。这就是 tools 这部分。再然后我们学护栏、学法律：在这个国家里什么是被允许的？你过马路要左右看，这是你的责任，不是司机的责任，尤其在繁忙的大路上更是这样。所以我们要讲护栏和法律。对。如果我们把这一整套都做了，再把这里的人换成一个模型——这个模型就类比于你从 OpenAI、Anthropic 或者 Google 那儿拿到的东西，模型只是那个智能层，它只是被训练来预测下一个词，现在还带了点推理能力，但你需要这一整套 harness，也就是脚手架（有人叫它 scaffolding），才能真正搭出一个能解决问题、能为你创造价值的东西。这套 harness 就叫 agent，然后我们用 n8n 这样的框架去搭这些 agent，对吧？再看，到目前为止我们搭过的、或者别人搭过的每一个 agent，里面都至少有这四样东西里的一个，或者……

[9:23] Mahesh

the good ones have all these things in it. So, let me show you like if you want to build these agents, what is the memory? Why you need this knowledge piece? What will happen if I don't add the knowledge? So, let's spend like maybe 10 minutes and understand the basics first before we get into how can you automate your world with these latest agent on how you become a builder PM. So, in my story, I would say the first step of becoming a builder PM is just getting into knowing the basics. So, if you look at this, maybe, so So is my just N8N, and what I'm doing here is maybe I start from scratch because what's the point of doing doing an Akash show without with with with nets. Let's do it without nets because that's the Trump he's acts without nets. So on the right hand side you just take an AI agent. And you search AI agent, you get an AI agent. So AI agent has a model which is the intelligence layer which I was just talking about. So now I can connecting an open AI chat model. You can pick any model you want. I will just save money because expensive models uh So I'm I'm going to be a little cheap here and pick the GP GPT 4.1 mini. So you can just pick that model. So now you got model. So this is the agent. This is like a little baby doesn't know anything about your world. And now it has intelligence. So okay, if it has intelligence, we should be ready to go. So I can ask it some questions like hi. What is What is What are neural networks? It can answer those questions. So now what happens is this goes to the agent. The agent calls the model. And one beautiful thing about this is that you can look at like what the message was sent, what the input and output were from this screen here. Okay? It gave me an input. It gave me because it's trained on these things. Okay then.

好的那些里面四样全有。那我给你演示一下，如果你想搭这些 agent——什么是 memory？你为什么需要这块知识？要是我不加知识会发生什么？所以咱们花个大概 10 分钟，先把基础搞懂，再进入怎么用这些最新的 agent 把你的世界自动化、你怎么成为 builder PM。在我的故事里，我会说成为 builder PM 的第一步，就是先去搞懂这些基础。所以你看这个——也许，这就是我的 n8n。我在这儿做的，是从零开始，因为在 Akash 的节目上不来点硬的、不带保护网地干，那有啥意思呢？咱们就不带保护网地来，因为他就是那种不带保护网上场的人。在右手边你就拖一个 AI agent，搜 AI agent，你就得到一个 AI agent。这个 AI agent 有一个模型，也就是我刚说的那个智能层。现在我可以接上一个 OpenAI 聊天模型。你想挑哪个模型都行，我就省点钱吧，因为贵的模型……所以我这儿稍微抠门一点，挑个 GPT-4.1 mini。你就挑那个模型。好，现在你有模型了，这就是 agent，它就像个什么都不懂你这个世界的小婴儿，现在它有智能了。好，既然它有智能了，那应该就能开干了吧。我可以问它点问题，比如：嗨，什么是……什么是神经网络？它能答上来。那现在发生的是：请求送到 agent，agent 调用模型。这里有个特别棒的点是，你能在这个界面上看到发出去的消息是什么、这块的输入和输出分别是什么。看到没？它给了我输入，给了我答案，因为它就是在这些东西上训练出来的。好，接着。

[11:29] Mahesh

But if I ask it very simple question like hey What is the What President Trump said about ending the conflict in Iran? What do you think it will answer that question or not?

但如果我问它一个特别简单的问题，比如：哎，Trump 总统关于结束伊朗冲突说了什么？你觉得这个问题它答得上来还是答不上来？

[11:49]

[laughter]

[笑声]

[11:50] Aakash

It's training data probably ended in like 2023. Let's see. Great answer. Let's see what happens. I ask this question and it says hey my knowledge cut off is June 2024 because you're cheap you're picking a cheap model. So that means that it doesn't have that knowledge so it can't answer the question. So maybe we need a tool for latest greatest things. So let's give it a tool and the tool can be Tavily which is a search tool. So this allows it to search the model to do things. So now I go and I say Tavily. I think it's A V I. Yep. This is the one. You got it. If you've been enjoying this episode today, [music] Mahesh's classic whiteboard style, his awesome labs and demos, then you will love Mahesh's cohort [music] based course. It has amongst the highest reviews on Maven and I challenge you to go check this yourself. His courses are always rated 4849. [music] That's because of the 13 years in big tech at Microsoft, Google, Amazon, Meta. He's actually built the things. He's not just talking about it and his course everybody I know, my mentees who have taken his course have had a positive opinion of it because it is so interactive. You get the first principles that you're sensing from the podcast today. The coolest thing about his new course, building open claws and with cloud code, is that you get a Mac mini. [music] So take advantage, use the discount code that I have below in the description so you get a discount on his course. Get your Mac mini. Better yet, get your company to pay for it. He's going to teach you not just how to become a builder PM, but he's going to be in the trenches with you. As he said, he will even give you AI PM [music] interview advice. So it's a all-in-one course to becoming an AI builder PM. Check that out in the link just below

它的训练数据大概到 2023 年就截止了。咱看看。好答案。看看会怎样。我把这个问题一问，它说：哎，我的知识截止到 2024 年 6 月，因为你抠门挑了个便宜模型。这就说明它没有那个知识，所以答不上来。那也许我们得给它配个工具，来搞定最新最热的东西。所以咱给它配个工具，这个工具可以是 Tavily，一个搜索工具。它能让模型去搜索、去干活。那我去输入 Tavily，我记得是 A-V-I……对，就是这个，你说对了。如果你喜欢今天这期节目、喜欢 Mahesh 经典的白板风格、他超棒的现场实验和演示，那你一定会爱上 Mahesh 那门基于 cohort 的课程。它在 Maven 上的评分名列前茅，我敢说你自己去看看也会服。他的课一向被打到 48、49 分。那是因为他在 Microsoft、Google、Amazon、Meta 这些大厂干了 13 年，是真刀真枪造过东西的人，不是光嘴上说说。我认识的每一个上过他课的人——我那些上过他课的徒弟——都对它评价很高，因为它太有互动性了，你能拿到今天在播客里隐约感受到的那种第一性原理。他新课最酷的一点——这门讲搭 OpenClaw、用 Claude Code 的课——是你能拿到一台 Mac mini。所以好好把握，用我放在简介里的折扣码，你报他的课就能打折，还能拿到你的 Mac mini。更妙的是，让你公司掏钱。他不光教你怎么成为 builder PM，还会和你一起泡在战壕里。就像他说的，他甚至会给你 AI PM 的面试建议。所以这是一门帮你成为 AI builder PM 的一站式课程，去简介下方的链接看看吧。

[13:37] Aakash

and back to today's episode. Today's episode is brought to you by Amplitude. Replays of mobile user engagement are critical to building better products and experiences, but many session replay tools don't capture the full picture. Some tools take screenshots every second leading to choppy replays and high storage costs from enormous capture sizes. Others use wireframes, but key moments go missing creating gaps in your understanding. Neither approach gives you a truly mobile experience. Amplitude does things differently. Their mobile replays capture the full experience. Every tap, every scroll, and every gesture with no lag and no performance hit. It's the most accurate way to understand mobile behavior. See the full story with Amplitude. And on query, so now I'm adding a tool. I have my account. It will go do the search. It will take the query. And where can it take the query? I can hard code the query or I can let the model define the query. So now I'm telling it that hey, let the intelligence layer based on what user is asking define what you want to search on internet. So now we connected a tool and let's go. So now I ask the same question again which is what is President Trump said about ending the conflict in Iran? And my spellings as you can see are not that great, but now it goes and now it's going to the search tool. It's doing some searching. Let's see what searching it is doing. Seems like it went to these website. It found information and then it went to a lot of website. It captured all the information and now it is able to answer the question. Which is from President Donald Trump stated that conflict in Iran could end at any moment based on his decision. He mentioned there is practically nothing left to target and all that. So a lot of excitement.

好，回到今天这期节目。今天这期由 Amplitude 赞助。移动端用户互动的回放对打造更好的产品和体验至关重要，但很多 session replay 工具没法捕捉到完整画面。有些工具每秒截一张图，导致回放卡顿，而且因为捕捉量巨大，存储成本很高。还有些用线框图，但关键时刻就丢了，让你的理解出现断层。这两种做法都没法给你真正的移动端体验。Amplitude 不一样，他们的移动回放能捕捉完整体验——每一次点按、每一次滑动、每一个手势，没有延迟、没有性能损耗，是理解移动端行为最精准的方式。用 Amplitude 看清完整故事。然后看查询这块——我现在加了个工具，我有账号了，它会去搜索，它会拿到 query。那它从哪儿拿这个 query 呢？我可以把 query 写死，也可以让模型自己来定这个 query。所以现在我告诉它：哎，让智能层根据用户问的内容，自己来决定要在互联网上搜什么。这样我们就接好了一个工具，开干。现在我再问一遍刚才那个问题——Trump 总统关于结束伊朗冲突说了什么？你也看到了我拼写不咋地，但现在它去调那个搜索工具了，正在搜，咱看看它在搜什么。看样子它去了这几个网站，找到了信息，然后又跑了一大堆网站，把所有信息都抓下来了，现在它就能回答这个问题了——它说 Donald Trump 总统表示，伊朗的冲突可以根据他的决定随时结束，他提到基本上已经没什么可打的目标了等等。所以挺让人兴奋的。

[15:18] Mahesh

I'm happy about that. So what is President Trump said? Okay. Then if I ask it a question which is hey What conflict am I talking about? You think that will work? Oh, I don't know. It doesn't have memory, right? Yeah, you you trained well.

我对这个挺满意。那么 Trump 总统说了啥？好。接着我要是问它一个问题：哎，我说的是哪场冲突？你觉得这个能行吗？哦，我不知道。它没有 memory 嘛，对吧？对，你学得挺到位。

[15:37]

[laughter]

[笑声]

[15:37] Mahesh

So if I ask it what what am what conflict? I don't see any previous mention of conflict in this conversation. Could you please provide more context or specify which conflict? So if you build an agent with the tool or intelligence, it will be a stupid agent because it doesn't have memory or it doesn't remember anything. Who wants to talk to a person who remember doesn't remember anything? So let's add a memory. So I will just add a simple memory which what it does is it takes a session ID and remembers last five conversations. Nice. So now you know what an agent is which is just a scaffolding, but the real work is happening in intelligence, memory, or tools. So if I do the same query now and it say hey what is the President Trump ending the conflict in Iran? And it will do the same thing, but this time you see it updates the memory. So it has put all the conversations, all the information in memory. And if I ask my question which is hey What is the conflict I'm talking about? It goes in. It fetches this from the memory. Now it doesn't call the tool because it sees like I'm looking for information. You are referring to the conflict in Iran specially in the military conflict on war situation begin around February 2026. So that's what it takes. But maybe maybe you have a larger context. Maybe you have contracts in your company and you want to ask questions on those contracts because conflict in Iran is not going to make me money. But if I do good contracts, it will make lot of money. So can I ask a question like hey What are the clauses or what are the payment terms impact of tariffs and war in Iran on payment payments impact payment term impacts of tariff and end of war on our contracts? Okay.

所以我要是问它——什么……什么冲突？它回答：我在这段对话里没看到之前提过任何冲突，能不能请你提供更多上下文，或者说清楚是哪场冲突？所以你要是只用工具和智能去搭一个 agent，那它会是个笨 agent，因为它没有 memory，啥都记不住。谁愿意跟一个什么都记不住的人聊天呢？那咱加个 memory。我就加一个简单的 memory，它做的事是：拿一个 session ID，记住最近五轮对话。不错。现在你就知道 agent 是什么了——它只是个脚手架，真正干活的是智能、memory 和 tools。所以我现在跑同一个查询，问它：哎，Trump 总统关于结束伊朗冲突说了什么？它会做一样的事，但这次你看，它更新了 memory，它把所有对话、所有信息都放进了 memory 里。然后我再问我的问题：哎，我说的是哪场冲突？它进去，从 memory 里把这个取出来，这次它不调那个工具了，因为它看出来我是在找信息。它说：你指的是伊朗的冲突，特别是大约在 2026 年 2 月前后开始的那场军事冲突、战争局势。所以这就是它的本事了。但也许……也许你有更大的上下文，也许你公司里有合同，你想就这些合同提问，因为伊朗的冲突又不能帮我赚钱，但我要是把合同搞好了，那能赚不少钱。所以我能不能问个问题，比如：哎，关税和伊朗战争对我们合同付款条款的影响有哪些条款、有哪些付款条款……关税和战争结束对我们合同付款条款的影响是什么？好。

[17:35] Mahesh

So I run a company and we have contracts all over the world. I want to know that. Will that work? I guess right now we haven't given it like a rag database, right? To our contracts. Yeah, I think I got the I got the best student here, right? Who has all the right prompts. So now it goes in and it says hey I don't know or I will just translate the payment impact on tariffs and war on Iran contracts generally can include several key factor, increase cost, payment delays, and all that, but it doesn't talk about my contracts. Yeah. So I have another lab that we cover in our uh course where you can actually create a knowledge base or database of your knowledge, how your world works. So this is important because the world is working perfectly how neural network works or world information, but your company information also need to be executed or needs to be built. So what this does is first one just you can upload contract. So I will just execute this workflow. Here I can upload my contracts. So I will just upload a contract from my company and what you will see is that it will go and create uh create Maybe we put this MSA in. So this is a master service agreement. Mhm. And what it does is once I go submit, it goes and creates the chunks. Now you can see there is a data loader. You can see insert data. And now you're learning a new concept how the data is getting into something that machines can understand because these agents or these machines don't understand like you and me understand in text. So now you're seeing something called an embedding model. The embedding model allows you to convert text into embeddings. You're also seeing something called a data loader. What is a data loader? Oh, it takes the type of data which is a binary which is

假设我经营一家公司，我们在全世界都有合同，我想知道这个。这能行吗？我猜现在我们还没给它配一个 rag 数据库，对吧？把我们的合同放进去。对，我这儿真是遇上最好的学生了，对吧？所有正确的 prompt 他全有。那现在它进去，它说：哎，我不知道，或者我就笼统翻一下——关税和伊朗战争对合同付款的影响，一般来说可以包含几个关键因素，成本上升、付款延迟之类的，但它讲的不是我的合同。对吧。所以我还有另一个 lab，是我们课程里讲到的，你可以在那里真正建起一个知识库，或者说一个属于你自己的、关于你这个世界怎么运转的数据库。这个很重要，因为这个世界本身运转得很完美——神经网络是怎么回事、世界的信息是怎样，这些都没问题，但你公司的信息也需要被处理、被搭进去才行。所以这个 lab 做的事是：第一步，你可以上传合同。那我就跑一下这个 workflow，在这儿我可以上传我的合同。我就从我公司上传一份合同，然后你会看到它去创建……去创建……要不咱把这份 MSA 放进去。这是一份主服务协议（master service agreement）。嗯。它做的事是：我一点提交，它就进去把内容切成 chunk。现在你能看到这儿有个 data loader，能看到插入数据。现在你正在学一个新概念——数据是怎么被弄成机器能看懂的形式的，因为这些 agent、这些机器没法像你我这样去理解文字。所以现在你看到一个叫 embedding 模型的东西，embedding 模型让你能把文本转成 embedding。你还看到一个叫 data loader 的东西，data loader 是什么？哦，它接收某种类型的数据，是一种二进制……

[19:40] Mahesh

where your input file, converts it into text, and then does simple text splitting. Okay? Maybe I can do custom splitting. Oh, so custom splitting you have to specify, but what's happening in simple one? Basically, it's taking 1,000 character with 200 character overlaps. So, it divides your whole file into 1,000 characters, and then it goes in and puts that into a database, or it calls the embedding model, which goes and creates this awesome database, which is a rag or retrieval augmented generation based database for you, which we have the same database where we are entering this information. And now, if I go here, and I ask the same question, which is I execute this workflow, and not this one. Give me this one. And I go I select this one, and I say execute workflow. Actually, not use this trigger. Just come here, and now I ask the same question which I was asking earlier. I can just say, "Hey, what is the value of contract? Does the payment term change due to war in an on tariffs? On tariffs based on our MSA contracts?" It goes in, it queries this tool, and this is the This is where you have put all your information. So, now it queries the tool, it extracts all the taxes information from your document that you provided, so your knowledge. And it says the document does not specify any provision, so I think we are good, whether it happens or not. And now this information comes from your knowledge. So, similar to this, we also discuss the next stage of this learning is multi-agent systems. So, single-agent systems are great, but if you want to do real checks or real things, this is a multi-agent system that we build. We use the same constructs that we just discussed. And in this one, you see that you can send an email. So, right now I can just send an email and ask this

把你的输入文件转换成文本，然后做简单的文本切分。明白吗？我也可以做自定义切分。哦，自定义切分你得自己指定参数，那简单切分里到底发生了什么呢？基本上就是按 1000 个字符、200 个字符重叠来切。它把整个文件切成 1000 字符一段，然后送进数据库，或者调用 embedding 模型——这个模型会帮你生成一个很棒的数据库，也就是基于 RAG（检索增强生成）的数据库，我们就是把这些信息录进同一个数据库的。现在，如果我到这儿来，问同样的问题，我执行这个 workflow，不是这个，给我那个。我选这个，然后说执行 workflow。其实别用这个触发器，直接到这儿来，现在我问刚才那个一样的问题。我可以直接问：「嘿，这份合同的金额是多少？根据我们的 MSA 合同，付款条款会因为战争或关税而改变吗？」它进去查询这个工具，这里就是你存放所有信息的地方。所以现在它查询这个工具，从你提供的文档里——也就是你的知识库里——提取出所有税务相关的信息。它回答说文档没有规定任何相关条款，所以我想我们没问题，不管这事发不发生。这个信息就是从你的知识库里来的。跟这个类似，我们再聊聊这套学习的下一个阶段，也就是多 agent（智能体）系统。单 agent 系统很好用，但如果你想做真正的核查、真正的事情，那就需要我们搭建的这种多 agent 系统了。我们用的还是刚才讨论过的那些构件。在这个系统里，你会看到你可以发邮件。所以现在我可以直接发一封邮件，去问

[21:59] Mahesh

information, the same information via an email. So, I have pasted that, or I have published that flow. And I can just ask like I ask my lawyers. Now I can send this request, and I say, "Hey, can you get me the risks in this contract before signing?" So, if you send me a contract, which I I I believe one day you should send me a contract, but

通过邮件去问同样的信息。我已经把这套流程粘贴好了，或者说发布好了。我可以像问律师那样去问它。现在我可以发出这个请求，我说：「嘿，能在签约之前帮我把这份合同里的风险找出来吗？」所以如果你发给我一份合同——我相信总有一天你应该给我发份合同的，不过

[22:22]

[laughter]

[笑声]

[22:23] Mahesh

uh but right now, let's say you send me a contract, which is an NDA. I want to Before signing, I will just send this email. And if I send this email, what happens is this email hits this provider, which is my Gmail. And if you look at it, I will get this review contract key terms for this contract. Okay, if I get that, then what happens is that if you I go to my workflow, I will show you an execution that in few seconds you will see that it will get a new running. So, it automatically triggers this. I need not to do anything. It's a published workflow. And now it's running. And when it is done, I will get a response, which says that, "Hey, you got these risks." Contract analysis report with all the risks in it. Nice. So, now you start first understanding all these things, which is, "Hey, what does these agents do, and how they work with your things behind the scenes?" So, these are the connectors. These are your agents. It has multi-agent system with playbooks or your database connected, and then it can do end-to-end work like humans do. And the last thing I would love to show in n8n world, which I think all of the builders PM should go build is This is great. All this is awesome. This is sending me risk. But what about evals? How are we going to evaluate these? Because these agents are not like us humans, which does a lot of self-evaluations, and you can't If they do a bad job, you are going to get fired. They are not going to lose their job. Even if they lose their job, you are It's not good for you. So, what is this is then then we create this idea of ground truths. So, what I have done here is I have taken this contract, and what I do is I say, "Hey, here are the terms that you need to look at. Here is the correct value, and whether

呃，不过现在，假设你发给我一份合同，是一份 NDA（保密协议）。在签约之前，我就直接发这封邮件。如果我发了这封邮件，会发生什么呢？这封邮件会打到这个服务提供方，也就是我的 Gmail。你看，我会收到这份「审阅本合同关键条款」的邮件。好，收到之后会发生什么呢？如果我去看我的 workflow，我给你演示一下，几秒钟内你就会看到有一个新的运行实例跑起来。它会自动触发，我什么都不用做。这是个已发布的 workflow。现在它就跑起来了。等它跑完，我会收到一个回复，说：「嘿，找到了这些风险。」一份合同分析报告，里面列出了所有风险。不错。所以现在你开始先理解这些东西，也就是「嘿，这些 agent 是干什么的，它们在幕后是怎么和你的东西配合工作的？」这些是连接器，这些是你的 agent。它有一套带 playbook 的多 agent 系统，连着你的数据库，然后它就能像人一样端到端地干活了。我想在 n8n 这个领域里展示的最后一件事——我觉得所有 builder PM 都该去搭一下——就是：这一切都很棒，它给我发了风险。但是评估（evals）呢？我们要怎么评估这些东西？因为这些 agent 不像我们人类会做大量自我评估。你要是干得不好，会被炒鱿鱼。它们不会丢工作。就算它们丢了工作，对你也不是好事。那这是什么呢？于是我们提出了 ground truth（标准答案）这个概念。我在这儿做的事情是，我拿了这份合同，然后我说：「嘿，这些是你需要关注的条款。这是正确的取值，以及

[24:20] Mahesh

there is a risk or not based on our playbook is also written from a real lawyer. And then, you can create a workflow like this, where you can run this normally in the daytime. What you do is during the day, you can have your people submitting these files. You can find all the risk, and the results are getting stored here. And in evening, what happens is you evaluate these using the judge or risk categories and modifications. And you get a fancy report, so let me just execute this. So, not the eval trigger, but let's just first create this data. So, if I go and execute this workflow, I can just upload a file, and on that file, what it's going to do is it's going to extract these terms. So, now it's going to find the risk in this contract, and submit our results according to the same key terms here. So, it runs in automation, and then it updates a file like a human will go find the risk, but what it does is it will just get the results first. So, now you will see once it is done, these values will be populated here. So, now it's just looking for governing law, justification, agreement terms. And once it is done, which is still executing, once it is done, you will see this file getting updated, and you will get the results like a human would have done. First, it's extracting these values from the contract, then checking whether risk or not. If risk, it will justify why this is a risk, and suggest a modification which allows you to take minimum risk if you sign this contract. This is what a lawyer does for you, by the way. They will look at your contract, compare it with some values, find whether risk or not, and then give you a justification why this is a risk and modification. They never justify the risk, by the way. So, now you see automatically that you

根据我们的 playbook 这里到底有没有风险」——这套 playbook 也是请真正的律师写的。然后你就可以搭一个像这样的 workflow，白天正常运行。白天你可以让你的人来提交这些文件，你能找出所有风险，结果就存在这里。到了晚上呢，你用这个评判器（judge），按风险类别和修改建议来评估这些结果，最后你会拿到一份很漂亮的报告。我来执行一下这个。先不用 eval 触发器，咱们先把这批数据造出来。如果我去执行这个 workflow，我可以上传一个文件，针对这个文件，它要做的就是提取这些条款。所以现在它要在这份合同里找风险，然后按照这里同样的关键条款提交我们的结果。它在自动化运行，然后它会更新一个文件——就像一个人去找风险一样，不过它做的是先把结果拿到。等它跑完，你就会看到这些值被填进这里。现在它正在找适用法律、理由说明、协议条款。等它跑完——现在还在执行——跑完之后你会看到这个文件被更新，你会拿到像人工做出来的那种结果。它先从合同里提取这些值，然后判断有没有风险。如果有风险，它会说明为什么这是个风险，并建议一个修改方案，让你在签这份合同时承担的风险最小。顺便说一句，这正是律师为你做的事。他们会看你的合同，跟某些标准值对比，判断有没有风险，然后给你一个理由说明为什么这是风险，再给个修改建议。顺便说一下，他们从不为风险辩护。所以现在你看到，它自动地

[26:22] Mahesh

can see the risk. And once you see these risk, then you can see the second part of this, which is you can run a quick eval on this by just going and changing it from form submission to eval. And now when I run it, what it does is it will go and run an automatic eval, and suggest comparing to my ground truth here, which is my real values, and suggest that how much of these risk is correct, how much of justification is what a human or these modification is what a human will accept as is, or they want to change. So, you can run these evals also, and n8n has this awesome tool. So, now you run this flow, and what it does is it will take it one row by one row, run your evaluations, and find whether the risk is good, and what is the quality of modification suggested by the AI. Mhm. It will continue doing it. But if you can go to evaluation tab, I can show you some of the previous evaluations. So, now you see that it goes through runs row by row, and then eventually it's telling you that, "Hey, you have a risk quality is good. You're able to detect 80% of risk correctly, but your modification quality, the suggestions AI is making, is only 30% as good as a human would have done. So, you have work to do." So, now you understand as a PM or as anybody who wants to survive this build wave, that it's not very hard to go with something very small, bring your context, make it multi-agent because the world is multi-agent, Mhm. and then run evaluations to make sure that before you deploy it, it's really good quality, so that you when it does a job or when it responds to these emails, it's it does a good job. So, I got my analysis which I sent earlier. So, it now it talks about all the findings. I need to set it up so that you can see it in a more The formatting could be formatting

就能看到这些风险。看到这些风险之后，你就能看到这套东西的第二部分，也就是你可以在上面跑一次快速评估——只要把它从「表单提交」改成「eval」就行。现在我一运行，它会去跑一次自动评估，跟我这里的 ground truth（也就是我的真实取值）做对比，告诉我这些风险里有多少是正确的，有多少理由说明是人类会认可的，这些修改建议有多少是人类会原样接受、又有多少是想改的。所以这些评估你也能跑，n8n 有这个很棒的工具。现在你运行这个流程，它会一行一行地处理，跑你的评估，判断风险找得对不对，以及 AI 给出的修改建议质量如何。嗯，它会接着往下跑。不过如果你去评估（evaluation）标签页，我可以给你看几个之前的评估结果。现在你看到它一行行地跑，最后它告诉你：「嘿，你的风险质量不错。你能正确识别出 80% 的风险，但你的修改质量——也就是 AI 给出的建议——只有人类水平的 30% 那么好。所以你还有活儿要干。」所以现在你作为一个 PM，或者作为任何想在这波 build 浪潮里活下来的人，就明白了：从很小的东西做起其实并不难，把你的上下文带进来，把它做成多 agent 的，因为这个世界本来就是多 agent 的——嗯——然后跑评估，确保在你部署它之前它的质量真的过关，这样它干活的时候、或者回复这些邮件的时候，就能干得漂亮。所以我拿到了刚才发出去的那份分析。现在它讲了所有的发现。我需要把它整理一下，让你能看得更……格式上其实可以

[28:29] Mahesh

correctly, but I just built it for you to show it quickly. But you see that the whole flow worked, and it does give you risk that this agreement shall be governed or constructed in law laws of state of Delaware without regard of conflict of law principle. Delaware is new neutral jurisdiction under criteria of NDA. So, it's talking about these things to you, and giving you all the details. Yeah. Okay. So, this is what it takes to get first your footing on ground to get started with your building journey. So, that's People talk about this when people talk about AI a lot, they talk about, "Hey, AI is You want to build a PM? Just start with Open Claw." My suggestion would be first understand what is the hardness looks like, what is the hardness made up. Ideally, understand how this model is built, what are neural networks, what are transformers. Uh and we talk about that in a way that anybody can understand. But then, understand the scaffolding at least, because this is where things will break even if you're using Claude code or the Claude code constructs like, you know, context, model compression, knowledge, memory. First, play with them and see what they are. And that's your first step in getting into building with AI. When does n8n fall short? When do you move beyond n8n? Yeah, so n8n is very like I think it's more like a tool which allows you to get to your first 10 customers. I think it's a very powerful tool, especially with webhook. When we did the last session, I showed you how can you create your backend in n8n without any code and connected to a lovable or v0 frontend and then build the whole solution where you can click a button and something happens in n8n and you can debug everything without writing a single line of code. I think it's very

弄得更规整些，不过我就是为了快速给你演示才临时搭的。但你看，整个流程跑通了，它确实给了你风险：本协议受特拉华州（Delaware）法律管辖和解释，不考虑法律冲突原则。在 NDA 的判定标准下，特拉华州是一个中立的司法管辖地。所以它就这样跟你讲这些东西，把所有细节都给你。是的。好。所以这就是你迈出第一步、在 build 之旅上站稳脚跟所需要的。当人们大谈 AI 的时候，他们会说：「嘿，你想当个 builder PM？直接从 OpenClaw 开始就行了。」我的建议是，你先理解这套底层硬核的东西长什么样、由什么构成。理想情况下，你得搞懂这个模型是怎么造出来的，什么是神经网络，什么是 transformer。呃，我们会用任何人都能听懂的方式来讲这些。但接下来，你至少要理解那套脚手架（scaffolding），因为就算你用的是 Claude Code 或者 Claude Code 的那些构件——比如上下文、模型压缩、知识、记忆——这里恰恰是东西会出问题的地方。先去玩玩它们，看看它们是什么。这就是你进入 AI build 世界的第一步。【Aakash】n8n 什么时候会不够用？什么时候你该抛开 n8n、往后走？【Mahesh】是的，n8n 非常……我觉得它更像是一个能帮你拿下头 10 个客户的工具。我觉得它是个很强大的工具，尤其是配上 webhook。上次那期，我给你演示过怎么不写一行代码就在 n8n 里搭出你的后端，然后接上 Lovable 或 v0 做的前端，再把整套方案搭起来——你点一个按钮，n8n 里就会发生点什么，而且你不用写一行代码就能把一切都调试出来。我觉得它在那方面非常

[30:16] Mahesh

powerful there. But then, if you want to iterate, put things in production, if you want three people to contribute to your code, if you want to have a test set, or want to put a container around it and put it into production, n8n doesn't support that. And if things fall short and the worst part of this is that there's no way for people to see the code and get to the code mode. Yeah.

强大。但接下来，如果你想做迭代、把东西放进生产环境，如果你想让三个人一起给你的代码做贡献，如果你想要一个测试集，或者想给它套一个容器再放进生产环境，n8n 就不支持这些了。而当它撑不住的时候，最糟糕的一点是，根本没法让人看到代码、进入代码模式（code mode）。对。

[30:40] Mahesh

Like we do this in a our cohorts and after people want to get to the next stage, which is hey, I've done this, but now I want to put this in code and share it with my team so that they can see if this is good quality, bad quality to take it to hundreds of user or thousands of users in the most efficient or latency optimized way, then n8n has no answer to those questions. So I think it just stops you beyond 10 customers, it's not the right tool I would recommend. And that's where I know what you want to get started with. Today's episode is brought to you by Jira Product Discovery. If you're like most product managers, you're probably in Jira tracking tickets and managing the backlog. But what about everything that happens before delivery? Jira Product Discovery helps you move your discovery, prioritization, and even roadmapping work out of spreadsheets and into a purpose-built tool designed for product teams. Capture insights, prioritize what matters, and create roadmaps you can easily tailor for any audience. And because it's built to work with Jira, everything stays connected from idea to delivery. Used by product teams at Canva, Deliveroo, and even The Economist, check out why and try it for free today at atlassian.com/product-discovery. That's a t l a s s i a n dot com {slash} product-discovery. Jira Product Discovery, build the right thing. Today's episode is brought to you by Naia One. In tech buying, speed is survival. How fast you can get a product in front of customers decides if you will win. If it takes you 9 months to buy one piece of tech, you're dead in the water. Right now, financial services are under pressure to get AI live. But in a regulated industry, the roadblocks are real. Naia One changes that. Their air-gapped, cloud-agnostic sandbox lets

就像我们在自己的训练营（cohort）里做的那样，到后面大家想进入下一个阶段，也就是「嘿，我把这个做出来了，但现在我想把它落成代码、分享给我的团队，让他们看看质量是好是坏，再以最高效、最优化延迟的方式把它推给成百上千的用户」，到这一步 n8n 就给不出答案了。所以我觉得它就是会把你卡在 10 个客户这个量级，超过这个量级它就不是我推荐的工具了。这也正好引出你想从哪儿开始上手的话题。今天这一期由 Jira Product Discovery 赞助播出。如果你跟大多数产品经理一样，你大概正在 Jira 里追踪工单、管理 backlog。但交付之前发生的那些事呢？Jira Product Discovery 能帮你把需求探索、优先级排序、甚至路线图规划的工作从电子表格里搬出来，放进一个专为产品团队打造的工具里。捕捉洞察、排定真正重要的事、做出你能轻松为任何受众量身定制的路线图。而且因为它就是为配合 Jira 而生的，从想法到交付，一切都保持连通。Canva、Deliveroo 甚至 The Economist 的产品团队都在用它，今天就去 atlassian.com/product-discovery 看看为什么，并免费试用。也就是 a-t-l-a-s-s-i-a-n 点 com 斜杠 product-discovery。Jira Product Discovery，做对的事。今天这一期还由 Naia One 赞助播出。在科技采购里，速度就是生存。你能多快把产品摆到客户面前，决定了你能不能赢。如果你买一件技术产品要花 9 个月，那你就已经没戏了。眼下，金融服务行业正承受着把 AI 上线的压力。但在一个强监管的行业里，障碍是实打实的。Naia One 改变了这一点。他们那套物理隔离（air-gapped）、不绑定任何云厂商的沙箱环境，能让你

[32:35] Mahesh

you find, test, and validate new AI tools much faster, from months to weeks, from stuck [snorts] to shipped. If you're ready to accelerate AI adoption, check out Naia One at naiaone.com/akash. That's n a y a o n e dot com {slash} a a k a s h. I hope you're enjoying today's episode. Are you interested in becoming an AI product manager making hundreds of thousands of dollars more joining OpenAI and Anthropic? Then you might want to do a course that I've taken myself, the AIPM certificate ran by OpenAI product leader Mik Dad Jaffer. If you use my code and my link, you get a special discount on this course. It is a course that I highly recommend. We have done a lot of collaborations together on things like AI product strategy. So check out our newsletter articles if you want to see the quality of the type of thinking you'll get. One of my frequent collaborators, Pavel Hearn, is the build labs leader. So you're going to live build an AI product with Pavel's feedback if you take this AIPM certificate. So be sure to check that out. Be sure to use my code and my link in order to get a special discount. And now back into today's episode. Yes, it's the hottest tool on the market, Claude Code. Please, Mahesh, show us when should we be using Claude Code? How should we be using it? Yeah, I think I think you should spend like good two weeks in n8n and beyond that you should move to Claude Code because I think what has happened is with Claude Code I'm and especially like in last, I would say, three to six months, Claude Code is there for more than a year now. But in last six months, there's just so much possibilities with Claude Code and Cowork to build things and put things in production for you as well as your team. It's the same tool chain which a person

用快得多的速度去发现、测试、验证新的 AI 工具，从几个月缩短到几周，从卡住不动到顺利上线。如果你准备好加速 AI 落地，去 naiaone.com/akash 看看 Naia One。也就是 n-a-y-a-o-n-e 点 com 斜杠 a-a-k-a-s-h。希望你正享受今天这一期节目。你有兴趣成为一名 AI 产品经理、多挣几十万美元、加入 OpenAI 和 Anthropic 这样的公司吗？那你可能会想上一门我自己也上过的课——由 OpenAI 产品负责人 Mik Dad Jaffer 开设的 AIPM 认证课程。如果你用我的优惠码和链接，能拿到这门课的专属折扣。这是一门我强烈推荐的课。我们在 AI 产品战略这类主题上有过很多合作，所以如果你想看看你能学到的那种思考的质量，去看看我们的 newsletter 文章吧。我经常合作的伙伴之一 Pavel Hearn 是 build labs 的负责人。所以如果你上这门 AIPM 认证课，你会在 Pavel 的反馈下实地搭建一个 AI 产品。一定要去看看。一定要用我的优惠码和链接来拿专属折扣。现在回到今天的节目。【Aakash】是的，市面上最火的工具，Claude Code。拜托了 Mahesh，给我们讲讲什么时候该用 Claude Code？该怎么用？【Mahesh】是的，我觉得你应该在 n8n 上扎扎实实花个两周左右，再往后就该转到 Claude Code 了。因为我觉得现在的情况是，有了 Claude Code——尤其是在最近，我会说，三到六个月里——Claude Code 其实已经出现一年多了。但在最近这六个月里，用 Claude Code 和 Cowork 去搭建东西、把东西推进生产、为你自己以及你的团队所用，可能性实在太多了。它是同一套工具链：一个完全没有编程经验的人

[34:18] Mahesh

who has no coding experience can use, like building with skills, creating their sub agents, hooks, and then schedule jobs to the people and then on the right hand side is people who know how to code, who has always been coding for rest of their lives. They also can build on top of what you provided. Yeah. So this idea of having a somebody to allocate or do work with and then this idea of code is combined in Claude Code. So I think it started with something that people wanted to do code with, but then this one thing they realize it is or we all realize is that if you can build something that codes well, it can do any task well. And that is what Claude Code is and that's what I think why it's the hottest tool in the town. And you talked at the beginning about needing to harness this moment. And Andre Karpathy talked about something changed in December 2025. What exactly changed? Yeah. So I think this is like for me also, right? I mean some contemplating this and I think still not have wrapped my head around it. What has changed is if you look at like last three years where I I was at Google and we were building AI, but I thought we are not building AI fast enough. So I left Google, I started on my own. I'm building this company where we thought we can go and automate back office and build AI or the benefits AI faster than what I could do at Google with my own agency. Obviously, Google is doing great with their agency. And if you look at all the companies in last three years, what they did with AI is they took these models that we had and if you take the examples of any company and I can put them what they do in this map, if you look at Gamma, Gamma did one thing. Gamma said, "Okay, I am going to go and connect these models with tools and make it very easy

也能用它——比如用 skill 来搭建、创建他们自己的 sub agent、hook，然后给团队成员派发定时任务（schedule jobs）；而另一边，右手边是那些会写代码、一辈子都在写代码的人。他们也能在你提供的基础之上继续往上搭。是的。所以「有一个人可以分配任务、一起干活」这个概念，跟「代码」这个概念，在 Claude Code 里结合起来了。所以我觉得它一开始是个大家想用来写代码的东西，但接着大家意识到的一件事是——或者说我们所有人都意识到的是——如果你能造出一个把代码写得很好的东西，那它任何任务都能干得很好。这就是 Claude Code 的本质，也是我觉得它为什么是当下最火的工具。你在开头讲到要抓住（harness）这个时刻。Andre Karpathy 也讲过，2025 年 12 月有什么东西变了。到底是什么变了？是的。我觉得这对我自己来说也是——我是说，我也一直在琢磨这件事，到现在我觉得自己也还没完全想明白。变了的是什么呢？如果你看过去这三年，我当时在 Google，我们在做 AI，但我觉得我们做 AI 的速度不够快。所以我离开了 Google，自己单干，做这家公司，我们觉得可以去自动化后台事务、把 AI 或者说 AI 的好处做得比我在 Google 凭自己那点能动性能做的还要快。当然，Google 用他们的能动性也做得很好。如果你看过去三年里所有这些公司用 AI 做了什么——他们拿了我们已有的这些模型——拿任何一家公司举例，我都能把它们干的事摆进这张图里。如果你看 Gamma，Gamma 做了一件事。Gamma 说：「好，我要去把这些模型和工具连起来，让你

[36:19] Mahesh

for you to create slides and then I will provide you connectors, connectors to PDFs, connectors to everything so that you can go and publish this in PPTs or Google Slides or just PDFs." They just did one job. And this was, I think, now at billion-dollar valuation. The second kind of company, so first kind of companies that you saw was these kind of companies which go and make the model and tool working and put lot of connectors in place. So Gamma is there. The second kind companies said, "Hey, we're going to take this model, harness it or domain-specific knowledge." So the Harveys of the world, the Legoras of the world went and said, "Hey, we're going to take this model and do the context engineering piece here and say, 'Hey, we will context to the lawyers and solve their problems.' And they become one to 10 billion-dollar companies in two years. Mhm. And some just went in and provided signals in memory well. And then there were third kind of companies which said, "Hey, can we provide you frameworks to build this?" Which is what you saw with Salesforce, Agent Force, Amazon Q. They said, "We will just help you build with all these things fast." Yeah. This was happening for last two years. You will see lot of companies just come out and every company was planning to do something like this. Then what happens is there is something, a breakthrough happens inside Anthropic where they are building just a tool for coding and they thought for doing coding, let's just do one thing which is build the agent loop. They called it agent loop. And the idea is it will build the context. It will take the actions, which is all the connectors and tools, and it will do the evaluations. And based on that, it can come back and keep doing that again and again. Mhm. So

能很轻松地做出幻灯片，然后我再给你提供各种连接器——连到 PDF 的、连到一切的连接器，这样你就能把它发布成 PPT、Google Slides，或者就是 PDF。」它们就只干了这一件事。而我觉得这家公司现在已经是十亿美元估值了。所以你看到的第一类公司就是这种：去把模型和工具跑通、再放进一大堆连接器。Gamma 就是这一类。第二类公司说：「嘿，我们要拿这个模型，给它注入领域专属的知识。」所以世界上那些 Harvey、Legora 这样的公司就去说：「嘿，我们要拿这个模型，在这儿做上下文工程这一块，说『嘿，我们把上下文喂给律师，帮他们解决问题』。」然后它们在两年里就成了估值一百亿到一百亿美元级别的公司。嗯。还有一些则是在记忆里把信号做好。接着出现了第三类公司，它们说：「嘿，我们能不能给你提供搭建这些东西的框架？」这就是你在 Salesforce、Agentforce、Amazon Q 身上看到的。它们说：「我们就帮你用所有这些东西快速搭起来。」是的。过去这两年一直在发生这样的事。你会看到很多公司冒出来，每家公司都打算去做点类似这样的事。然后发生的事情是，Anthropic 内部出现了某种突破——他们当时只是在做一个用来写代码的工具，他们想，为了写代码，咱们就做一件事，就是把 agent loop（智能体循环）搭出来。他们管它叫 agent loop。这个想法是：它会构建上下文，会采取行动——也就是所有那些连接器和工具——然后它会做评估。基于评估结果，它能再回过头来，一遍又一遍地不停这么干。嗯。所以

[38:29] Mahesh

now, what happens is if you were a context company, Harvey, Legora, if you are an action company, Gamma, if you are an eval company or you did evals to make sure that your actions and all this is good, all this is coming part of this product called Claude Code. Yeah. And now people realize that it's not only coding specific, you can actually do work with it. So then they release these plugins for legal, marketing, and sales. And what happens is they are able to do a better job at context management, taking actions, and evaluations. Why? Because for context, there is another unblock in this world, which is what we thought I thought of that like the best thing that will happen with these models, but I thought maybe it is the browser where things will happen. But where things happened is this idea of computer control. And we all became normal because we thought these are coding tools. So we gave them access to two things. One, your file system. And second, your bash commands. This is how the whole world works. This is how if you're an engineer, you can control your computer. And if you can control your computer and your browser, and you can have access to your file system, you can do all the context management. And with bash, you can do all the actions. And third, it has the browser control as well. So now, you have all the action powers. And on top of it, they have evaluations, which is they are saying, "Hey, I can do the length length checks, do rule-based checks. I can go and do UI element checks in HTML. And I can do what I was showing you, which is these LLMs as a judge. And all this I am going to do. So you don't need a third-party provider. You don't need Harvey. You don't need somebody to come in and do it for you. If you take co-work and if you

现在发生的事情是这样：如果你是一家做 context 的公司，比如 Harvey、Legora；如果你是一家做 action 的公司，比如 Gamma；如果你是一家做 eval 的公司，或者你做 eval 来确保你的 action 这些东西都没问题——所有这些能力，现在都变成了 Claude Code 这个产品的一部分。对。现在大家意识到，它不只是用来写代码的，你真的可以用它来干活。于是他们就发布了面向法律、市场营销、销售的这些插件。结果就是，他们能把 context 管理、采取 action、做 evaluation 这几件事做得更好。为什么？因为在 context 这一块，这个世界上还有另一个被解锁的东西，就是我当初以为——我曾经觉得这些模型能带来的最好的东西，我以为可能会发生在浏览器上。但真正发生突破的地方，是这个叫 computer control（控制电脑）的概念。而我们都觉得这很正常，因为我们以为这些是写代码的工具。所以我们给了它们两样东西的访问权限：第一，你的文件系统；第二，你的 bash 命令。整个世界就是这么运转的。如果你是工程师，你就是这样控制你的电脑的。而如果你能控制你的电脑、你的浏览器，还能访问你的文件系统，你就能做所有的 context 管理。再加上 bash，你就能执行所有的 action。第三，它还有浏览器控制能力。所以现在，你拥有了所有的执行能力。在这之上，他们还有 evaluation，他们说：「嘿，我可以做长度检查、做基于规则的检查，我可以去 HTML 里做 UI 元素的检查，我还可以做我刚才给你演示的那种——用 LLM 当评委（LLM as a judge）。所有这些我都能做。所以你不需要第三方供应商了，你不需要 Harvey 了，你不需要别人来帮你做了。如果你用 co-work，如果你

[40:45] Mahesh

use my tested skills, you can do it yourself. So this is first time that whole intelligence layer that we were talking about, which was flawed, actually became the whole harness provider as well. So this is Anthropic entering into everybody's lunch and saying, "Hey, we don't we need your the space of this also." And that two to three billion what you're seeing in Claude code is basically coming right now from these valuations which these people thought that they can go solve for each customers. But now they can't solve because everything is mostly three things. So once you have this, then on top of it, you can connect a UI layer. Or these three things, the how they are able to teach the world. And obviously, you have done a great job with all your podcast and teaching the world these skills on top of it, which is that anything that you want to do can be a skills. Anything. And skills are powered by these action-taking sub-agents, which can be triggered by something called hooks. And now you can schedule jobs. With these people learning these on top of these computer control, which is powered by this context loop. And by the way, there is a bottom layer to this, which is these models like Opus 4.6, have grown and able to do long-horizon jobs. What is that? They are trained in a way that they can run for three to six hours without breaking. This is was not there. If you looked at the matter benchmark, the last six months ago, the longest horizon job they can do is three minutes. And this has gone exponentially in last six months. So you put a long-horizon job, you give me access to your bash file file system, and then the users can create these skills naturally in English language. You have unlocked what the human potential was locked inside these

用我测试过的这些 skill，你自己就能搞定。所以这是第一次，我们刚才聊的那个原本有缺陷的整个智能层（intelligence layer），实际上也变成了整个 harness 的提供方。所以这就是 Anthropic 杀进了所有人的饭碗，说：「嘿，这块空间我们也要了。」而你看到 Claude Code 那两三亿（美元）的收入，基本上现在就是来自这些 evaluation——这些人原本以为他们可以替每个客户去解决这些问题，但现在他们解决不了了，因为一切归根结底基本就是这三件事。所以一旦你有了这个，在它之上，你还可以接一个 UI 层。这三件事，是他们教会全世界的方式。当然，你（Aakash）也做得很棒，用你的播客教全世界这些 skill，在这之上还有一点：任何你想做的事情都可以变成一个 skill。任何事情。而 skill 是由这些会采取 action 的 sub-agent（子智能体）驱动的，这些 sub-agent 可以被一个叫 hooks 的东西触发。现在你还可以调度任务（schedule jobs）。这些人在这套 computer control 之上学会了这些东西，而 computer control 又是由这个 context 循环驱动的。顺便说一句，这底下还有一层，就是像 Opus 4.6 这样的模型——它们已经成长到能干长周期（long-horizon）的活儿了。这是什么意思？它们被训练成可以连续跑三到六个小时不崩。这以前是没有的。如果你看那个 benchmark，六个月前，它们能干的最长周期任务也就三分钟。而过去六个月这个能力是指数级增长的。所以你扔给它一个长周期任务，你给我访问你的 bash 和文件系统的权限，然后用户就能用英语很自然地创建这些 skill，你就解锁了那些原本被锁住的人类潜能，那些原本被锁在这些

[43:03] Mahesh

coding or some software or some software provider will go and do it for you. Now it's just a skill. And if it is a good skill, and if you know your skill, and if you have your craft, you can put it in a skill, which can be paralyzed with sub-agents and triggered with events like hooks. And then whole this operating system is available to everybody. And I think that's what happened with Claude code. Claude code people realized that they need not to do the hard piece of which I was showing you with rag, what the chunk size needs to be. And this rag, we do it in our labs with an attend where I show people that it's so hard to do with the right retrieval. And check this right retrieval with evals. So you have to do agentic rag, then you have to do graph rag. All that now is responsibility of Claude code. It goes, it gets the right context. If the context becomes too large, it compacts the context. All that is happening for you. You need not to code for it. You just pay for it. And you pay by by like some 20 bucks or 200 bucks, which Harvey, by the way, was charging you $10,000 to do. So that's what every lawyer got now. That's what every law firm has now. And they are looking and saying, "Maybe we don't need tools." And that's how the benefits of AI are getting diffused into economy with Claude code, then co-work. And maybe we can talk about open Claude next. But that's what changed for me six months ago when I actually took hold of Claude code with the latest greatest model. I tried Claude code, by the way, and I was not very impressed with it for doing things which are beyond coding. And I was like, "Yeah, it's a good tool. Yeah, but Cursor does the same thing. I have GitHub Copilot, which I got for free because I'm Microsoft, you know. So

写代码的人、或者某个软件、或者某个软件供应商里、要他们去帮你做的事情。现在它就只是一个 skill 而已。如果这是个好 skill，如果你懂你自己的 skill，如果你有自己的手艺，你就可以把它做成一个 skill，这个 skill 可以用 sub-agent 并行化，可以用 hooks 这样的事件来触发。然后整个操作系统就对所有人开放了。我觉得这就是 Claude Code 身上发生的事。Claude Code 让大家意识到，他们不需要去做那些苦活——我刚才用 RAG 给你演示的、chunk size 该设多大这种。而这种 RAG，我们在实验室里会带学员一起做，我给大家演示过，做对正确的检索（retrieval）有多难。还要用 eval 去验证这个检索对不对。所以你得做 agentic RAG，然后还得做 graph RAG。所有这些现在都是 Claude Code 的责任了。它会去帮你拿到正确的 context；如果 context 太大了，它会帮你压缩 context（compact）。所有这些都在替你做。你不需要为它写代码，你只需要付钱。而你付的钱大概也就 20 块或者 200 块，而这些事，Harvey 顺便说一句以前要收你一万美元才肯做。所以这就是现在每个律师都拥有的东西，每家律所现在都有的东西。他们看着这个就说：「也许我们不需要工具了。」AI 的红利就是这样通过 Claude Code、然后 co-work 扩散到整个经济里的。也许我们接下来可以聊聊 OpenClaw。但对我来说，六个月前真正改变的就是这个——那时候我真正上手了用最新最强模型的 Claude Code。顺便说，我以前也试过 Claude Code，对于用它做写代码以外的事情，我当时其实不太满意。我当时想：「嗯，是个好工具。但 Cursor 也能干一样的事。我还有 GitHub Copilot，因为我是微软的人所以免费拿到的，你懂的。所以」

[44:54] Mahesh

uh but then, three months ago, I went to Claude code once Opus 4.6 came out. And now it's a superpower. I don't know its limitations. It does anything that I want it to do. And that's why when I talk about it, people say that I'm scaring them to take my courses. My life is good without my courses, by the way. I'm just telling you that anything like when I looked at ChatGPT, I was very uh excited about it. When I looked at Loveable, I was excited. That was the second moment for me that yes, front-end is a solved problem.

呃但接着，三个月前，等 Opus 4.6 出来之后，我又回去用了 Claude Code。现在它就是个超能力。我都不知道它的极限在哪儿。我想让它干什么它就能干什么。所以当我谈起它的时候，有人说我是在吓唬大家来报我的课。顺便说一句，没有我那些课我的日子也过得挺好。我只是告诉你——就像我当初看到 ChatGPT 的时候，我特别兴奋；我看到 Lovable 的时候也很兴奋，那是我的第二个时刻，就是觉得：是的，前端是个已经被解决的问题了。

[45:26]

[snorts]

[吸鼻子声]

[45:26] Mahesh

And then I looked at Claude code, and it seems like everything is a solved problem. It can replace anybody. And this is an open challenge I have given to anybody that if you can do a job, just tell me the job, and this agent will do it better than you 365 days, 24 hours. And that's what we are playing with now, right? So that's the world you live in. And this is moving each model this long-horizon jobs are shifting to more and more jobs. These context action evals are getting better inside. So you need not to worry about it. And more and more people are sharing their skills, their agents, and their scaffolding. And this is the new scaffolding layer, which is a very thin layer in English, rather than figuring out all rag, all of this tool calling, all of that putting the guardrails. Obviously, guardrail is still your responsibility, but but there is like at least you can rely on these companies to not screw up because that's the only thing they need to go solve next. So that's my take on the question, right? Which is, "Hey, what has changed for all of us?" What has changed for all of us is that if you know your skills, you can build it in Claude code and delegate that work to agents. Can you show us this in action? What does it look like? Oh. You you like action so much, so let me show you that also. And again, right?

然后我看到了 Claude Code，感觉好像一切都是已经被解决的问题了。它可以取代任何人。我有一个公开的挑战，对任何人都成立：只要你能干某个活儿，你就告诉我那是什么活儿，这个 agent 一年 365 天、一天 24 小时都会比你干得更好。这就是我们现在在玩的东西，对吧？所以这就是你身处的世界。而且这件事还在随着每一代模型不断推进——这些长周期任务正在覆盖越来越多的工种。这些 context、action、eval 在内部越做越好。所以你不用操心它们了。而且越来越多的人在分享他们的 skill、他们的 agent、他们的 scaffolding（脚手架）。这就是新的 scaffolding 层，它是一个非常薄的、用英语写的层，而不用你去搞懂所有的 RAG、所有的工具调用（tool calling）、所有那些设护栏（guardrails）的事。当然，护栏仍然是你的责任，但至少你可以信任这些公司不会搞砸，因为那是他们接下来唯一需要去解决的事。所以这就是我对这个问题的看法，对吧——也就是「嘿，对我们所有人来说什么变了？」对我们所有人来说，变了的是：如果你懂你的 skill，你就可以在 Claude Code 里把它造出来，然后把那份工作委派给 agent 去做。（Aakash）你能现场给我们演示一下吗？它看起来是什么样的？哦。你这么喜欢 action，那我也给你演示一下这个。还是那句话，对吧？

[46:52]

[clears throat]

[清嗓子]

[46:52] Mahesh

So there is one job I do, right? Because I started my own company. I have like at this point 20 people working with us. And once you do that, you figure out that most of your time goes into reviews. So what I build for myself here is that what I was able to do with just uh I just thought maybe I just do this. Let me share. As as you can see, this is my screen. Uh this is my uh this is my co-work. I have bunch of stuff going on. What people send me all the time is that they will send me a PRD for review. So now I have given the review job to my agent. How it works? If you send me a PRD, I will just go and say, "Hey." Or I can just start a new task. And I will say, "Hey." And I will select my folder. So this is co-work. Same like Claude code, you can provide it a context. So I will select my review con- context. I will change my model. Opus 4.6 is too expensive. So let's do Sonnet. So

有这么一件活儿是我经常干的，对吧？因为我自己开了公司，到现在大概有 20 个人跟我一起干。一旦你管这么多人，你就会发现你大部分时间都花在了 review（评审）上。所以我给自己造了这么个东西——其实就是我随口一想：也许我就把这个做出来吧。我分享一下屏幕。你们能看到，这是我的屏幕。呃这是我的 co-work。我这上面有一堆东西在跑。大家老是发给我的，是各种要我 review 的 PRD。所以现在我把 review 这份活儿交给了我的 agent。它怎么工作呢？如果你发给我一个 PRD，我就会直接说「嘿」，或者我可以直接开一个新任务。我会说「嘿」，然后选我的文件夹。这就是 co-work，跟 Claude Code 一样，你可以给它提供 context。所以我会选我的 review context。我会改一下模型。Opus 4.6 太贵了，所以咱们用 Sonnet 吧。所以

[47:53]

[laughter]

[笑声]

[47:54] Mahesh

So I will go to Sonnet because I'm paying them a lot, so they're like, "Let's eat his token with Opus and give him the best results." No, Sonnet is good enough. Uh so I go to the review folder. And now I can upload a file and say, "Hey, can you just put comments on it?" So if the good model or before this, if you were doing something like this, what will happen is you have to go and create So now you can get So this is a product two-pager. We are just building a new product. This is live. And what happens here is that I can go and upload this file and then say, "Hey, can you use our checklist and review this file, put comments as I would have done?" So now, if you look at this, this is some sophisticated tool that you need because now it need my checklist, which I already provided in these reviews. But first, it needs to build the context. It need to understand. But now I you need not to build the query engine, the query to knowledge mapping. Now it automatically finds out the skill which I have built. It's my PRD review skill. And then it goes to my reviews instructions. And then it's reading my checklist. And based on my checklist which I have provided to it, it's going to review this file and then upload or put comments in it like I would have done. As I showed you earlier, I could have put this whole agent inside Slack. And somebody could have just asked my review, and this would have given a review in five minutes and saved me bunch of time on basic things which people forget, of course, because they are busy with their lives. So that's the first.

所以我会切到 Sonnet，因为我付了他们很多钱，他们大概想着「让他用 Opus 多烧点 token，给他最好的结果」。不用，Sonnet 已经足够好了。呃所以我进到 review 文件夹。现在我可以上传一个文件，说：「嘿，你能帮我把批注（comments）加上去吗？」如果是个好模型——或者说在这之前，如果你想干这种事，会发生什么呢？你得自己去搭建……所以现在你能直接拿到……这是一份产品的两页纸（two-pager），我们正在做一个新产品。这是真实的、上线的东西。这里发生的是，我可以上传这个文件，然后说：「嘿，你能用我们的 checklist（清单）review 这个文件，像我会做的那样把批注加上去吗？」现在如果你看，这本来需要一套相当复杂的工具，因为它需要我的 checklist——我之前已经在这些 review 里提供过了。但首先它得先建立 context，它得理解。但现在你不用自己去搭查询引擎、搭「查询到知识」的映射了。现在它自动就找到了我造的那个 skill，是我的 PRD review skill。然后它会去读我的 review 指令，然后它在读我的 checklist。基于我提供给它的这个 checklist，它就会去 review 这个文件，然后像我会做的那样在里面上传或者加上批注。就像我刚才给你演示的，我本来还可以把整个 agent 塞进 Slack 里，然后随便谁都可以来找我 review，它就会在五分钟内给出一份 review，帮我在那些基础的事情上省下一堆时间——那些大家容易忘的事，当然了，因为他们都忙着过自己的日子嘛。所以这是第一个。

[49:38] Aakash

Can [laughter] [laughter] [laughter] we look inside the PRD review

我们能[笑声][笑声][笑声]看看这个 PRD review 的内部吗

[52:25] Mahesh

Mode is missing. Add section explaining defensible advantage. What are the steps that a doc or big four consulting firms from building this? AI failure modes are unaddressed. What happens when attribution is wrong? How do you handle misclassification of AI versus human work?" So, not only like some wishy-washy, the real good comments which I will put in. So, that's your first step. Maybe you got it right, and maybe you didn't. Maybe So, if you go and build an agent, then this should be it. If you build a chat chat, then this should be it. But, then you will go and I look at this and I say, "Yeah, they did the good job, but I don't like this whole section-wise." So, then I will put another comment, and I will say, "Hey, you know what? This thing is This is more market specific. How can you make it?" So, now I am adding my comments, and I'm saying it looks for PR fact for you write in that format. This means that uh heading problem and solution section without question answering here. Question and answer format. So, seems like format was picked was not taken, and that was not in our So, I added a new comment here. Okay. So, the beautiful part earlier was that uh yes, it did the job, but obviously it might miss few things. And I today I have a very different angle. One day I go and comment everything. One day I don't comment anything. So, today I want to push AI. The idea is that if AI does our work, what are we going to do? We're going to do and push push it even further. So, now I'm doing my job, which is I'm saying, "Hey, you go this comment." And then I look at this, and I say, "Why this problem is hard? General observability tool don't understand. Workflow multi-step, multi-tool. Adoption is invisible. This is too broad." So, I can say, "Hey, this is too broad.

缺少了 Mode（模式）这一节。补充一节解释一下你的护城河（defensible advantage）是什么。一家四大咨询公司要造出这个东西需要哪些步骤？AI 的失败模式（failure modes）没有被处理。归因（attribution）错了会怎么样？你怎么处理把 AI 的工作误判成人类工作、或者反过来的情况？」你看，不是那种含糊其辞的，而是真正高质量的、我自己会写的那种批注。所以，这是你的第一步。也许你一下就做对了，也许没做对。也许……所以如果你去造一个 agent，那就应该是这样。如果你造一个 chat（对话），那就应该是这样。但接着你会去……我看着这个，我说：「嗯，他们干得不错，但我不喜欢这种整个按章节来的方式。」所以我就再加一条批注，我会说：「嘿，你知道吗？这个东西……这个更偏特定市场。你能不能把它改成……」所以现在我在加我自己的批注，我说它应该照着 PR FAQ 那个格式来给你写。意思是呃要用「问题—解决方案」的标题，而不是在这里用一问一答（question answering）的方式。要用问答（Q&A）格式。所以看起来它选的格式没被采纳，而那本来也不在我们的……所以我在这儿加了一条新批注。好。所以早先那个漂亮的地方在于，呃是的，它把活儿干了，但它显然可能漏掉几样东西。而我今天的角度很不一样。有一天我会去把所有东西都批注一遍，有一天我什么都不批注。所以今天我想推一推 AI。我的想法是：如果 AI 替我们干活了，那我们要干嘛？我们就去把它再往前推一推、再推一推。所以现在我在干我的活儿，也就是我说：「嘿，你看这条批注。」然后我看着这个，我说：「为什么这个问题很难？通用的可观测性（observability）工具理解不了。Workflow 是多步骤、多工具的。采用（adoption）是看不见的。这个太宽泛了。」所以我可以说：「嘿，这个太宽泛了。」

[54:32] Mahesh

This is too broad. Make it pointed. Make it point to human story as what happens to CXOs when they can't see impact of their AI investments." Great. So, now I have put two comments. Okay. So, So, this was uh output file that you got, and now you got this whole thing. Then, we what I build on this is that I said, "Great. Now, AI does the review work for me because I have these skills. I have put my instructions in the Claude code, and it does the work." But, if you have somebody who is working for you, or if you work in real environments, you want it to learn from it every day. So, for that, I build another skill, another sub-agent, which is What it does is it goes in, and I have scheduled them to go and check for my review comments. So, what this one does is it goes and runs every 2 hours or every 30 minutes, and checks what are the comments I am adding. It automatically is scheduled. It looks in the same folder where in the reviews I'm putting all my things. And it goes, and then what it does is it creates this file called learner.md. And it learns from my patterns. Oh. So, now if you look, it's updating this file, which is a learner.md, which Claude code did not provide it. This is what as a human I added. So, it goes and put all my files. First, every day whenever this job is done, because I'm a PM, I do a good job of organizing things because I need to evaluate because I understand these concepts. So, every time a job is done, it creates a folder here which says who did the job, what was the job about, when it happened. And inside that, it creates this that this is the checklist I used. This was the input document. This is the output document. And this is what the user modified document looks like. So, it create all the artifacts. So, if

「这个太宽泛了。把它写具体一点。让它指向一个具体的人的故事——当那些 CXO（高管）看不到他们 AI 投资的影响时，会发生什么。」很好。所以现在我加了两条批注。好。所以，这就是你之前拿到的那个输出文件，现在你拿到了这一整套东西。然后我在这上面又造了一个东西，我说：「好，现在 AI 替我做 review 这份活儿了，因为我有这些 skill，我把我的指令放进了 Claude Code，它就把活儿干了。」但是，如果你手下有个员工，或者你身处真实的工作环境里，你会希望它每天都能从中学习。所以为了这个，我又造了另一个 skill、另一个 sub-agent，它干的事是：它会进去——我把它们调度成会去检查我的 review 批注。所以这个 sub-agent 干的事是，它每 2 小时或者每 30 分钟跑一次，去检查我都加了哪些批注。它是自动调度的。它会去看同一个文件夹，就是我在 review 里放所有东西的那个文件夹。然后它会去创建一个叫 learner.md 的文件。它从我的模式（pattern）里学习。哦。所以现在你看，它在更新这个文件，这个 learner.md 文件，这是 Claude Code 没有提供的，是我作为人类自己加上去的。所以它会去把我所有的文件都放进去。首先，每天每当这份活儿干完了——因为我是 PM，我很擅长把东西整理得井井有条，因为我需要去评估，因为我理解这些概念——所以每次一份活儿干完了，它就在这儿创建一个文件夹，里面写着是谁干的活、这活儿是关于什么的、什么时候干的。在那个文件夹里面，它还会创建这些：这是我用的 checklist，这是输入文档，这是输出文档，而这是用户改过之后的文档长什么样。所以它会把所有这些产物（artifact）都创建出来。所以如果

[56:53] Mahesh

you want to go back, check, you have all the data that you need to debug or see what happened in this story. And then, each 30 minutes, Claude goes and compares these two files, and creates this learner.md file which says, "Hey, I looked at this job folder, and user added these comments. And with that, I have these new learnings. Yep. What Claude What I got right, what I missed, and these are the checklist that I want to go and update in the checklist. But, it doesn't go and update it right away. I have another skill that sees the patterns here. And if it finds it for 5 days, it sends me an email and says, "Hey, I want to update your checklist. I have seen that you revised this many times. You have updated this same comment five times today, or five times in past week. Do you want me to update the checklist? Here is the updated checklist." And I quickly review it, and then my new checklist comes into life. So, now I have How does that look? Yeah. So, that is the PRD review checklist. This checklist, and then you will see versions of that this checklist. Mhm. Which is getting stored above this whole thing. Mhm. So, there is a master version, and then there there is a version that is in works. Mhm. So, now if you look at it, what I'm trying to do is Okay, because I'm a human, right? And I have to push beyond what these folks have already solved for. So, what they have solved for is you can create these skills, but they are pretty static. This is what basically I think all the people are trying to solve for. Let me just fix the camera first. Okay. So, if you look at [clears throat] it, what I'm trying to solve is Let's start here.

你想回头去查，你就有了所有需要的数据，可以去 debug、可以去看这件事到底发生了什么。然后，每 30 分钟，Claude 会去对比这两个文件，然后创建这个 learner.md 文件，它会说：「嘿，我看了这个任务文件夹，用户加了这些批注。基于这些，我有了这些新的学习收获。对。Claude……我哪些做对了、哪些漏掉了，以及这些是我想去更新进 checklist 的条目。」但它不会马上就去更新。我还有另一个 skill，它会观察这里的模式。如果它连着 5 天都发现了同一个模式，它就会给我发一封邮件，说：「嘿，我想更新一下你的 checklist。我注意到你把这条改了很多次。你今天把同一条批注更新了五次，或者过去一周里更新了五次。你想让我更新 checklist 吗？这是更新后的 checklist。」我快速 review 一下，然后我的新 checklist 就生效了。所以现在我有了……（Aakash）那个看起来是什么样的？对。这就是 PRD review checklist。这个 checklist，然后你会看到这个 checklist 的各个版本。嗯。它被存在这整个东西的上面。嗯。所以有一个主版本（master），然后还有一个正在改的版本。嗯。所以现在如果你看它，我想做的是……好吧，因为我是个人类嘛，对吧？我得往前推，超越这些人已经替我解决掉的东西。他们已经解决的是：你可以创建这些 skill，但它们是相当静态的。这基本上就是我觉得所有人都在试图解决的问题。我先把摄像头调一下。好。所以如果你看[清嗓子]它，我想解决的是……咱们从这儿开始。

[58:51]

[clears throat]

[清嗓子]

[58:52] Mahesh

Okay. Then, what I'm trying to solve for is that yes, this agent loop can solve the world hunger. But, it can't solve my hunger. So, what I have done is I, as Mahesh, call this agent loop. We take the learnings of this, which is the learners.md. And then I have my agent, which sits on top of it, which is inside machine. So, this is Mahesh. Box is AI. And it goes and checks the work done by this agent loop in those folders, and then keep updating my comments or my work every day and creates this learner.md which updates the checklist which is an input to this loop. Not without my feedback, with my feedback. Mhm. So now I have created a continuous learning loop. So now every day

好。那么，我想解决的是这个：是的，这个 agent 循环可以解决世界饥饿问题（拯救全世界），但它解决不了我自己的饥饿（我的个性化需求）。所以我做的是，我，作为 Mahesh，调用这个 agent 循环。我们把这个循环的学习收获，也就是这个 learner.md，拿出来。然后我有我自己的 agent，它坐在这之上，它在那台机器里。所以这是 Mahesh，方框是 AI。它会去检查这个 agent 循环在那些文件夹里干完的活儿，然后每天不断更新我的批注、更新我的工作，再创建这个 learner.md，由它去更新那个 checklist，而这个 checklist 又是这个循环的输入。不是不经过我的反馈，而是带着我的反馈。嗯。所以现在我创建了一个持续学习（continuous learning）的循环。所以现在每天

[59:54]

[clears throat and cough]

[清嗓子和咳嗽声]

[59:58] Mahesh

Now what happens is every day when you come in and I use this system, my reviewer is going to be better than Akash reviewer or anybody else's reviewer. Yeah. Or Claude code reviewer because it has learned from my best practices. And everybody who's improving this file, maybe five people comment on this file. My organization has a different soul. We care about different things. And now all this can also be contributed if we build on top of the agentic loop. Which is bringing our own context and this is the idea of continuous learners which I have like first step today. But there is more obviously we all can follow. But that's not coming inside the loop. This is outside the loop that I'm trying to build with this learner checklist and feedback. Mhm. And once you do that, you can adapt these continuously. Now you have a continuous learner like humans. So you and me go to any company like this podcast and last time to this time I'm improvising based on the feedback you gave me and based on the comments all of you gave me. Same way these agents are also learning now based on what you provide as feedbacks, which is very natural. We are not asking you thumbs up and thumbs down. We're just asking you to do the job that you gave it to us. And if you would added to that job, the next time we will do better. So keep doing your jobs, keep these agents along with you and they will learn your world and every day they will get better and push you to push you to even get better at your job. And that I think is a dream for all people like if you have a craft and if somebody is there to learn from your your craft from you and then push you further. I think that's the dream humans always had. So we got a preview into PRD comments, but what else should PMs be using Claude code to do?

现在发生的事情是，每天当你来用这套系统的时候，我的 reviewer 会比 Aakash 的 reviewer、或者其他任何人的 reviewer 都更好。对。或者比 Claude Code 的 reviewer 更好，因为它从我的最佳实践里学到了东西。每一个改进这个文件的人——也许有五个人在这个文件上加批注——我的组织有它不同的灵魂，我们在乎不同的东西。现在如果我们在这个 agentic 循环之上去搭建，所有这些也都可以被贡献进来。这就是带上你自己的 context 的理念，也就是 continuous learner（持续学习者）的想法，这只是我今天迈出的第一步。但显然后面还有更多，我们都可以跟进。但那些不在这个循环里面，那是在循环外面、我正在用这个 learner、checklist 加反馈去搭建的东西。嗯。一旦你做到这一点，你就能持续地适应这些。现在你就有了一个像人类一样持续学习的东西。所以你和我去任何一家公司，比如来上这个播客，从上次到这次，我都在根据你给我的反馈、根据你们所有人给我的批注来即兴调整、改进。同样的，这些 agent 现在也在根据你提供的反馈来学习，这非常自然。我们不是让你点赞或者点踩，我们只是让你去干你交给它的那份活儿。如果你在那份活儿上有所补充，下一次我们就会干得更好。所以你就继续干你的活儿吧，带着这些 agent 一起，它们会学会你的世界，每天都会变得更好，并且推着你、推着你在你的工作上做得更好。我觉得这对所有人来说都是个梦想——如果你有一门手艺，又有个东西在那里向你学习你的手艺，然后再推着你往前走。我觉得这是人类一直以来的梦想。（Aakash）所以我们看到了 PRD 批注的预览，但 PM 们还应该用 Claude Code 来做些什么呢？

[1:01:55] Mahesh

Uh I think first is like do your first thing which is like you started in any 10, then you created your first agent which you were able to do a normal job that you're doing every day. Like for me PRD reviews was the thing. For you maybe writing PRDs is the thing. And you do that first job, but then after that, I think the genie is out of the box. Let me show you what PMs do and what we help PMs do in our cohort. So what happens is once you understand all these cool things that they exist then what can you do with these? You can just go. So this is our first lab that you can use which says hey, creating agents in Claude code, we talk about all the basic things that you need to do and this one just does competitive analysis. So this is building your own competitive analyzer. So it will go research the web and gives you insight. So this is a lab which we give our creating sub agents. Here you create sub agents which basically what they are doing is we have lot of sub agents now inside Claude code and they will go and look at different competition competitors you have and what's the insider news, what's the outsider news and generate a report for you. Oh, that's not enough. Then this one allows you to create mocks. So okay, you can create once you do your user research market research, you can create PRDs in Claude code, but you can create mocks and visualization which earlier you were waiting for your design teams to do. Oh, but that's not fun maybe. Maybe you can clone your mocks from the source and then modify the mocks to build an end-to-end prototype of your product. So now with here we build a whole product from the mocks. So we take these screens and we modify them and build an end-to-end working prototype that you can publish and customers can

嗯，我觉得第一步就是先把你的第一件事做出来，比如你从 n8n 起步，然后造出你的第一个 agent，让它能帮你干一件你每天都在做的常规活儿。对我来说就是 PRD 评审，对你来说可能就是写 PRD。先把这第一件活干成，但做完之后，我觉得就一发不可收拾了。我给你看看我们 cohort 里 PM 都在做什么、我们帮 PM 做什么。事情是这样的，一旦你理解了这些很酷的东西都存在，那你能拿它们干什么呢？你就可以直接上手。这是我们第一个实验课（lab），你可以用它，它讲的是在 Claude Code 里创建 agent，我们会讲到所有你需要掌握的基础，而这一个做的就是竞品分析。所以这是在搭你自己的竞品分析器，它会去网上做调研，给你输出洞察。这就是我们给学员的「创建 sub agent」的实验课。在这里你创建 sub agent，它们做的事情就是——我们现在在 Claude Code 里有很多 sub agent，它们会去看你面对的各种竞品，看看有什么内部消息、有什么外部消息，然后给你生成一份报告。哦，这还不够。然后这一个让你能创建原型图（mock）。所以，你做完用户调研、市场调研之后，可以在 Claude Code 里写 PRD，但你还能创建 mock 和可视化，而这些以前你得等设计团队来做。哦，但这也许不够有意思。也许你可以从源头把 mock 克隆下来，再改这些 mock，搭出你产品端到端的原型。所以现在我们在这里直接从 mock 把整个产品搭出来。我们拿这些界面，改一改，搭出一个端到端能跑的原型，你可以发布出去，让客户去——

[1:03:44] Mahesh

feel, touch and give you feedback. Not in mocks. As a real product. And then you can analyze data. So you can see who's using, how they are using, where things are failing. So then we give you this lab which allows you to not only just give you data, but also analyze this data and create fancy dashboards which shows you how many contracts have been analyzed, what is the processing time, what is the average rating, compliance rate, bug reports, all. So your app gathers the data and now you are creating these dashboards. So everything which used to take you almost two to three months first to write the PRD to get to mocks, from mocks to a real working prototype from there to getting customers and seeing the signals all that is getting squeezed with this Claude code and all this you can do with Claude code. And we give you labs for that. And the labs are public. So for all Akash audience, we will even do one hour free sessions. Whatever it takes. Because the mission is to make sure that everybody can build in this new age because there is lot of spend on these. This technology is very expensive and if we can't build, if we can't diffuse the benefits of this technology to economy, we all are going to fail. So with that at least I'm going to do my part and Akash is doing his part by allowing all of us this platform to spread whatever we have learned or whatever we have seen out there. So Claude code, it's obviously great. It's a session based power tool. What are the limitations of Claude code and when should PMs be thinking about using open Claude? Oh, another one. So this is a this is an amazing session. We start with like chatbots, then we go to any 10, then we go to Claude code and now the most beautiful thing called open Claude. So as we talked about this agentic loop,

——感受、触摸，给你反馈。不是在 mock 上，而是当作一个真实的产品。然后你还可以分析数据。你能看到谁在用、怎么用、哪里出问题。所以我们又给你这个实验课，它不只是把数据给你，还能分析这些数据、生成漂亮的 dashboard，告诉你分析了多少份合同、处理时长是多少、平均评分、合规率、bug 报告，全都有。所以你的 app 收集数据，现在你来生成这些 dashboard。所以以前从写 PRD 到出 mock、从 mock 到能跑的真实原型、再到拉来客户看信号，整套流程几乎要花你两三个月，现在全被 Claude Code 压缩了，这一切你都能用 Claude Code 做。我们也为此提供实验课，而且这些实验课是公开的。所以对所有 Aakash 的观众，我们甚至会做一小时的免费课程，不管要付出什么。因为我们的使命就是确保人人都能在这个新时代里搞创造，因为在这些东西上的投入很大，这项技术非常昂贵，如果我们造不出东西、如果我们没法把这项技术的红利扩散到整个经济里，我们就都会失败。所以至少我会尽我的一份力，Aakash 也在尽他的一份力，他提供这个平台，让我们能把学到的、看到的东西传播出去。所以 Claude Code 显然很棒，它是一个基于会话（session）的强力工具。

[1:05:40] Aakash

right? What happened? This is the another exciting thing that happened in December and generally what happens is November, December people get time to actually do things and like sit down and not just run through these errands of life. So what happened with that is that Peter who is a developer from Australia by the way. And all these [snorts] things needs to come from developers and not from big companies. That's the one pattern you are seeing that the big big breakthrough came from ChatGPT which is like an open AI one team, few people launched something in and then they could see amazing thing happen. Similarly lovable a team outside US just built something and became overnight sensation. And then you saw open Claude. So what is open Claude? So open Claude said, hey now this agentic loop is open. Anybody can build on it. So initially what they said is open Claude because this agentic loop is coming in agent agent SDK. So I can write the same tool like co-work or Claude code. Anybody can write that. And the beautiful part is that it can connect to any model. So when Claude build it or Anthropic build it, the agentic SDK was open that it doesn't only work with Claude models. It can work with any models. So what he did is he said this loop is great, but people are having problems connecting to different channels. So first layer what he did is he connected to these channels which is WhatsApp signal and slack Telegram and everything. Yeah. And at this point like hundreds of these. And then he created this gateway which automatically opens a port and make sure that these are good. So he did the hard work of taking the formats and making sure that all can be processed. And then what I was passing through my email when I get this example, you were

那 Claude Code 有什么局限，PM 又该在什么时候考虑用 OpenClaw 呢？

[1:07:50] Mahesh

sending this email, but I was passing it to the model. He takes all these inputs or puts this agent SDK and then says hey, this is the intelligence layer. And now anybody sending me these messages here I will process it here and then come back and you can connect your tools you can connect your models same way. And all this is coming as one thing called open Claude. Mhm. And by the way, I'm going to open source all of this and give everybody free access to it. So it's not tied to any company. It's out there. Anybody can audit it. It's not something. So this is the first version of open Claude. So he said I'm going to open source Claude which by the way like today yesterday the code got leaked. But this is the code code getting leaked in a very nice way. And now if you look at [clears throat] it because it's developed by a developer and because it has two things that Claude code didn't had in that time. One, it allowed everybody to just go and delegate work to it. So as you were talking about like the session, the idea of Claude code was that it was built for developers and developer does iterations. But here there was not iterations. Here what they did is they did delegate the work. So you You delegate the work. I will go do do do the work, and when the work is done, I will come back to you. And you need not to bother about it because I'm directly coming to you through your channels, and you need not to be in terminal with me and go back and forth. So, one idea is the delegation idea. Second idea is this idea of a shell or this sandbox. So, instead of you giving me permissions on every file, every folder, why can't you just install me on a machine? So, install me on a Mac Mini. That's why Mac Mini is uh out of out of orders or 3 weeks delay now

哦，又一个好问题。这真是一场精彩的对谈。我们从聊天机器人讲起，然后到 n8n，再到 Claude Code，现在到了那个最美妙的东西——OpenClaw。所以就像我们聊过的这个 agentic loop，对吧？发生了什么？这是又一件在去年十二月发生的激动人心的事，通常呢，每年十一月、十二月人们才有时间真正坐下来做点东西，而不只是被生活的琐事推着跑。所以那时候发生的事是，有个叫 Peter 的人——顺便说一句他是个来自澳大利亚的开发者。而所有这些东西都得来自开发者，而不是大公司。这就是你看到的一个规律：大的突破来自 ChatGPT，那是 OpenAI 的一个团队、几个人发布了点东西，然后他们就看到神奇的事情发生了。同样，Lovable 是个美国之外的团队，造了点东西，一夜爆红。然后你又看到了 OpenClaw。那 OpenClaw 是什么？OpenClaw 说，嘿，现在这个 agentic loop 是开放的了，任何人都能在上面搭东西。一开始他们说叫 OpenClaw，是因为这个 agentic loop 是放在 agent SDK 里出来的。所以我可以写一个跟 co-work 或 Claude Code 一样的工具，谁都能写。而美妙的地方在于，它能接到任何模型。所以当 Claude——或者说 Anthropic——做这个的时候，这个 agentic SDK 是开放的，它不只跟 Claude 模型搭配，它能跟任何模型搭配。所以他做的事是，他说这个 loop 很棒，但大家在连接不同渠道时遇到了麻烦。所以第一层他做的，是把它接到这些渠道上，也就是 WhatsApp、Signal、Slack、Telegram 这些。是的，到这个程度已经有上百个了。然后他造了这个 gateway，自动开一个端口，确保这些都没问题。所以他做了苦活——把各种格式拿过来，确保全都能处理。然后我之前是通过邮件——我收到这个例子的时候，你是——

[1:10:00] Mahesh

because if I can install this, this is the new operating system. These These tools became the UI or the mouse clickable things that you can assign things, and now this whole compute works for you. And the third unlock was you can connect any model and even open-source models. And you are no more tied to the limits that Claude could basically everybody is hitting every day. So, now you can tied it to any open-source model. So, that is Open Claw for you. So, Open Claw one So, this is the template that you will see everywhere now. So, this template is going to be the next operating system if my predictions have been right, and I'm predicting all these trends all along. By the way, I've said agents will be great in 2023. Then I said, "Hey, you know what? Uh we're going to go to multi-agent orchestration in 2025." And this year I've been saying that this pattern that you have seen here will be copied all over again and again and again, and everybody will build on top of these. So, now this is the new agentic layer or this is the new agent definition. And now you can give the work, and the work is getting done by the agent, and you get the output back. So, you can measure input and output rather than measuring tools, evaluations, and all. So, that's Open Claw for you, and you can just see it in action as well because I know you will just say, "Hey, can you show me in action?" So, You have to transition.

——你是在发这封邮件，但我是把它传给了模型。它把所有这些输入接进来，放进这个 agent SDK，然后说，嘿，这就是智能层。现在任何人在这里给我发消息，我就在这里处理，然后回过头来，而你可以接你的工具、接你的模型，方式都一样。这一切作为一个东西呈现出来，就叫 OpenClaw。嗯。顺便说一句，我会把这一切全都开源，让所有人免费用。所以它不绑定任何公司，它就摆在那儿，任何人都能审计它，它不是什么黑箱。所以这是 OpenClaw 的第一个版本。他说我要把它开源——顺便说，今天、昨天代码被泄露了，但这是一种很漂亮的代码泄露方式。现在如果你看它，因为它是一个开发者做的，因为它有两样 Claude Code 当时没有的东西。第一，它让所有人都能直接把活儿委派（delegate）给它。所以正如你说的那个 session，Claude Code 的理念是它是为开发者造的，而开发者是要做迭代的。但在这里没有迭代。这里他们做的是委派工作。所以你把活儿派出去，我去把活儿干完，干完之后我回来找你，你完全不用操心，因为我是直接通过你的渠道来找你的，你不用待在终端里跟我来回折腾。所以一个理念是委派的理念。第二个理念，是这个 shell 或者说沙箱（sandbox）的理念。所以，与其让你给我每个文件、每个文件夹的权限，为什么你不干脆把我装在一台机器上呢？把我装在一台 Mac Mini 上。这就是为什么 Mac Mini 现在缺货、要延迟三周才能发货——

[1:11:40] Mahesh

Now I know, right? I've done it enough times with you. So, what I have done is I have By the way, I have my Mac Mini here also, but for today, what I have done is I have done simple installation because it's hard to just show you my Mac Mini, and there's just a lot of things going on that which I can't share a lot about. But what There is another way to put Mac Mini or Open Claw beyond Mac Mini is this UTM. So, on your Mac machine, you can create this new VM using this tool. And I have my whole setup. By the way, we have labs for these. So, you can see that this is my session. You can see my overview. You can see which channels I have connected. I have connected only WhatsApp as a channel. And now you can see my usage of it, which is pretty okay because this is not what I use daily. The daily is in the Mac Mini. And then you can set up cron jobs, which is like scheduled jobs. You can define your agent skill nodes here also. So, how How can you delegate work? Okay. I can just go in here. I can go to my WhatsApp, and I can just start chatting with it. I can say, "Hey, do a deep research on what are What is agentic loop and long horizon jobs capability and how they can impact software market. Give me full report." By the way, I can say all this through my command line as well. So, it goes in now. What it is doing is it's going to process this request and give me results. And this is me sending a message on WhatsApp to a friend or to an agent on the other side of the world, and you see that it says, "Hey, here is the agentic loop introduction, landscape, emergence of autonomous software agents, increased demand, companies leveraging." And this is just vanilla. I've not even put my skills and everything that I have put on others. But now you can just

——因为如果我能把这个装上，这就是新的操作系统。这些工具变成了 UI、变成了你能点的鼠标，你把活儿派给它，现在这整台算力就为你工作了。第三个解锁是，你能接任何模型，甚至开源模型。你再也不被那些 Claude 的限制捆住了——基本上每个人每天都在撞那个上限。所以现在你可以把它接到任何开源模型上。这就是 OpenClaw。所以 OpenClaw 这一个——这就是你以后到处都会看到的那个模板。这个模板，如果我的预测一直没错的话，将会是下一个操作系统，而我一直在预测这些趋势。顺便说一句，我在 2023 年就说过 agent 会很厉害。然后我说，嘿，你知道吗，2025 年我们会走向多 agent 编排（multi-agent orchestration）。而今年我一直在说，你在这里看到的这个模式会被一遍又一遍、一遍又一遍地复制，所有人都会在这之上搭东西。所以现在这就是新的 agentic 层，或者说这就是新的 agent 定义。现在你把活儿交出去，活儿由 agent 来完成，你拿回输出。所以你可以衡量输入和输出，而不是去衡量工具、评估那些东西。所以这就是 OpenClaw，你也可以直接看它跑起来——因为我知道你肯定会说，嘿，能不能给我演示一下实际运行？所以，你得切换一下。

[1:13:52] Mahesh

put a channel, and in 30 minutes we give you labs as well. So, for this also, we have set up labs where you can go and create your whole Just give me 1 second. I think those are Okay, this is coming down. Open Claw Open Claw. So, here also, we have labs for you where you can go and set it up with WhatsApp, but you can also automate all of your world using this lab where you can connect your Gmail, which I have connected to my Mac Mini, which I couldn't show because just there is so much personal stuff. But you can connect and make it your personal assistant. You can let it manage your calendar by step-by-step following this guide. Second thing what we have done on this one is that you can go and create a whole discovery process or whole autonomous developer for you, which goes and scans your GitHub and then fixes the bugs that are P2 or P3 for you. And then send a PR request to your development team. So, now you are becoming If they are becoming a threat to you, you are becoming a threat to them. So, you can build the whole lab where it scans your GitHub. First, the agent goes and does tests for your UI or a new feature, then files the bugs, and then instead of assigning these bugs to developers, it goes and try to fix the bugs as well, and then send a pull request for final approval to engineering because they still control the code. But you can build all that in Open Claw. I'm also building a mini PM in Open Claw, which will do all the PM's jobs. But first I thought maybe I should build the dev because that's what is scarcity for me. Maybe the engineers are building mini PM. So, that's Open Claw. If you can take these kind of a structure, you can set and assign work to it through the channels that you are already familiar with, and then you can have this idea of

现在我懂了，对吧？我跟你做过够多次了。所以我做的是——顺便说一句，我这儿也有我的 Mac Mini，但今天呢，我做的是一个简单的安装，因为我的 Mac Mini 不太好直接展示，上面有太多事情在跑，很多我没法多说。但是——把 Mac Mini 或者说把 OpenClaw 放到 Mac Mini 之外的另一种方式，是用这个 UTM。所以在你的 Mac 机器上，你可以用这个工具创建一个新的虚拟机（VM）。我整套环境都在这儿。顺便说，我们有这些的实验课。所以你能看到这是我的 session，你能看到我的概览，能看到我接了哪些渠道。我只接了 WhatsApp 这一个渠道。现在你能看到我对它的使用量，还挺一般的，因为这不是我日常用的，日常是在 Mac Mini 上。然后你可以设置 cron job，也就是定时任务。你也可以在这里定义你的 agent 技能节点。所以，怎么委派工作呢？好，我就进到这里，进到我的 WhatsApp，直接开始跟它聊。我可以说，嘿，帮我做个 deep research，主题是什么是 agentic loop、长周期任务（long horizon job）的能力，以及它们会怎样冲击软件市场，给我一份完整报告。顺便说，这些我也可以通过命令行来说。所以它现在进去了，它正在做的是处理这个请求、给我结果。这就是我在 WhatsApp 上给一个朋友、或者说给地球另一端的一个 agent 发消息，然后你看它说，嘿，这是 agentic loop 的介绍、行业格局、自主软件 agent 的兴起、需求增长、各公司如何利用。而这还只是最原始的版本，我连我放在别处的那些技能都还没加上去呢。但现在你就可以——

[1:16:06] Mahesh

this whole sandboxing or controlling a whole machine to itself. And you are just giving it that work and then permissions to go access it to do your jobs. So, you've been a PM at all these big tech companies like Google. Let's be real, right? Google isn't going to allow you to just give your company access to an Open Claw. How should a PM at a big company mitigate security concerns? How should they be using this latest technology? Yeah, I know I 100% agree. I think the idea is not like Open Claw is not a technology like it's not a product. It's a pattern for me. It's a pattern on how these agents can be useful within agentic loop. And they will copy the pattern and offer you in a sandbox way, which can be inside their anti-gravity or inside their Gmail workspace or on GCP. So, in GCP, if you ask somebody today and say, "Hey, my Kubernetes cluster is down." I can't debug that today, but with this pattern now, this message will be sent to their sandbox VM, which will be running Open Claw or some version of similar pattern. And now it will go and simulate first to their Kubernetes cluster, try to make the same deployment, and then debug it. First, reproduce the whole problem. And that's what people do. Like as humans, we will first try to reproduce that problem. We will try to make the same cluster do it. But it These agents can't do it. Single loops can't do it. But now we can do it because we have the full control on a full machine, and now the agent can create, reproduce your problem, suggest a solution, try the solution, and then come back to you. And that pattern on that VM is fully controlled by Google. And as a user, all you are seeing is I provided a solution to your problem. But you don't know that I tested it, but now I can test it.

——整套沙箱化、或者说让它掌控一整台机器的能力。你只是把那个活儿交给它，再给它权限去访问、去干你的活儿。

Aakash：你在 Google 这些大厂都当过 PM。咱们说实话，对吧，Google 是不会允许你随便把公司的访问权给一个 OpenClaw 的。大公司的 PM 该怎么化解这些安全顾虑？他们又该怎么用上这项最新的技术？

Mahesh：是的，我知道，我百分之百同意。我觉得关键在于，OpenClaw 它不是一项技术，它不是一个产品，对我来说它是一种模式（pattern）。它是一种关于这些 agent 如何在 agentic loop 里发挥作用的模式。各家会把这个模式复制过去，以沙箱化的方式提供给你——可能是放在他们的 Antigravity 里、放在他们的 Gmail workspace 里、或者放在 GCP 上。所以在 GCP 上，今天如果你去问某个人，说，嘿，我的 Kubernetes 集群挂了。我今天没法去 debug 它，但有了这个模式，现在这条消息会被发到他们的沙箱 VM 里，那上面跑着 OpenClaw 或者某个类似模式的版本。然后它会先去模拟你那个 Kubernetes 集群，试着做同样的部署，再去 debug。先把整个问题复现出来。这就是人会做的事。作为人类，我们会先试着复现那个问题，我们会试着搭一个同样的集群来跑。但是——这些 agent 以前做不到，单个 loop 做不到。但现在我们能做了，因为我们对一整台机器有完全的掌控，现在 agent 可以创建、复现你的问题，提出一个方案，试这个方案，然后再回来找你。而那台 VM 上的那套模式是完全由 Google 掌控的。作为用户，你看到的全部就是「我给你的问题提供了一个方案」。但你不知道我已经测试过它了——而现在我能测试它了。

[1:18:07] Mahesh

And that I think is what I am excited about. Obviously, there are challenges all around the security layer, which I think you already talked about in your podcasts earlier where a lot of skills, lot of attacks around those. But I think once you sandbox it, which I think is the next big thing. Now, what is left, right? We have the agent loop. We have the whole pattern. Then what is left? I think the ability to sandbox these agents in a controlled way, that's an unsolved problem, and Google will solve it, and I think Open AI is solving it, and that's what you saw with MyKano also. By the way, this idea of Open Claw is not new like MyKano, the the personal agent company gave you a VM, and they run their code. And if you look at it last time I looked at it, I can actually log in to that VM and start doing web browsing on their VM or do whatever I wanted to that VM. So, this was like this idea of enabling lot of possibilities was always there. Open Claw just made it so popular or so famous because it was open-source. So, Google will bring it to their companies in their sandboxes and solve end-to-end problems which humans used to do on their laptops and then dismantle that based on each query or each problem they solve. So, can you put it all together what we've learned today? How do we organize this knowledge? What drawers do we put it in? Where does basic knowledge about chat GPT and innate and agents, Claude code and open claw, how does it all come together to create that builder PM? Yeah, that's a great question. So, you say Mayur you said a lot of things. How can I just have a plan for it? So, I think first two to three weeks just understand the basics. Without that you won't be able to leverage or even understand it will become overwhelming

而这一点，我觉得就是让我兴奋的地方。当然，整个安全层四处都有挑战——我想你之前在播客里也聊过，有很多技能、很多针对这些的攻击。但我觉得一旦你把它沙箱化，我认为那就是下一个大事。现在还剩下什么呢，对吧？我们有了 agent loop，我们有了整套模式。那还剩什么？我觉得就是以一种受控的方式把这些 agent 沙箱化的能力，那是一个尚未解决的问题，Google 会解决它，我觉得 OpenAI 也在解决它，这也是你在 Manus（MyKano）那儿看到的。顺便说，OpenClaw 这个理念并不新鲜，就像 Manus 那个私人 agent 公司给你一个 VM、在上面跑他们的代码一样。如果你看的话，我上次看的时候，我其实可以登进那个 VM，在他们的 VM 上做网页浏览、或者在那个 VM 上想干什么都行。所以这种开启大量可能性的理念，一直都存在。OpenClaw 只是因为开源而让它变得这么火、这么出名。所以 Google 会把它带进他们公司的沙箱里，端到端地解决那些过去人类要在自己笔记本上做的问题，然后根据每一次查询、每一个解决的问题把它拆掉。

Aakash：那你能不能把我们今天学的东西串起来？我们怎么把这些知识组织起来？该把它们放进哪些抽屉里？关于 ChatGPT、n8n 和 agent、Claude Code 和 OpenClaw 的基础知识，它们怎么汇到一起，造出那个 builder PM？

[1:19:59] Mahesh

once you reach the open claw stage. So, I spend with my people like people who join our cohort I spend like first six weeks with them explaining them what the models is, what the intelligence or knowledge is, how these tools actually are working. So, spend that first three weeks. Then get to Claude code or co-work and automate your world. Which is whatever you do now agents should be doing and you should be building systems which basically allow agents to continuously learn or follow your patterns. Which was two things in my example today, my checklist and my learner. And then this human in loop pattern which was update everything every night but keep me in loop. So, that the second thing I would love you to build as a second stage. And third thing I would love you to spend another month on is just understand in and out of open claw and see how can you have one thing in your lifetime, in your job that you can just give it to a machine and the machine does the work and give you results somewhere else. And the whole machine will be controlled by this agent give permissions left and right. Make sure that you're not giving permissions to your world just create a separate world for this and then see if you can delegate work and get it done. And once you have done these three three things then just read obviously your newsletter or take any product and see is it a variant of open claw or agent loop or it is something that is starting from scratch like model knowledge. And then you will be able to see what are the possibilities that exist out there and which possibilities work for your company, for your feature, for your product that you're going to build next. And that's your next three weeks. So, this is like three weeks of first four weeks then three weeks then two weeks.

嗯，这是个好问题。你会说，Mahesh 你说了一大堆，我怎么能就有个计划呢？所以我觉得，头两三周就先把基础搞明白。没有这一步你就没法用好它，甚至没法理解它，等你到了 OpenClaw 这个阶段，会变得手足无措。所以我跟我的学员——加入我们 cohort 的人——我头六周就花在给他们讲模型是什么、智能或知识是什么、这些工具实际是怎么运作的。所以先花头三周。然后到 Claude Code 或 co-work，把你的世界自动化。也就是你现在做的任何事，都应该由 agent 来做，而你应该去搭那些系统，让 agent 能持续学习、或者遵循你的模式。在我今天的例子里这就是两样东西：我的 checklist 和我的 learner。还有这个「human in loop」（人在回路）的模式——每晚更新一切，但把我留在回路里。所以作为第二个阶段，这是我希望你搭的第二样东西。第三样我希望你再花一个月去搞的，就是把 OpenClaw 里里外外都摸透，看看你能不能找到这一生中、你工作里的某一件事，是你可以直接交给一台机器、机器把活儿干完、再把结果送到别处给你的。而整台机器都由这个 agent 掌控，给它各种权限。务必别把你（真实）世界的权限给它，而是给这个单独造一个世界，然后看看你能不能把活儿委派出去、把它干成。等你把这三件事都做完了，那就去读——当然，读你的 newsletter，或者拿任何一个产品来看，看它是 OpenClaw 或 agent loop 的某种变体，还是某个从零开始的东西、比如从模型知识起步的。然后你就能看到外面存在哪些可能性，以及哪些可能性适合你的公司、你的功能、你下一个要做的产品。那就是你接下来的三周。所以这就像是：头四周、然后三周、然后两周。

[1:21:57] Aakash

So, you're looking at nine to 10 weeks of a good work through of building with AI or becoming a builder PM. So, you became a builder PM. Now you're trying to interview for the role. How has the PM interview changed with AI? What should PMs expect in the new AI PM interviews? Yeah, I think there is one thing that I would love to love love to share with everybody because I'm doing a lot of research on this and a lot of people are coming to me every day. My calendar is booked for 15 minutes calls. I do four of those with our cohort members. Uh just 15 minutes because they have interviews and I don't charge for it if you've done our cohort uh because I just want to help and I want to stay updated as well on what's happening. So, let me tell you like three things that's are happening. One is this idea especially for level five level six AI roles that idea of doing normal product sense is gone. You will be given a problem and you will be asked to solve it either in a case study or during your interview. That is becoming a pattern. And in that people are trying to see, "Hey, do you understand where we stand, how the world is working or you are stuck in some past like six months ago or a one year ago past? So, are you the person who is going to take us to the new world or drag us with the old world making old decisions?" So, that's what the first thing I'm going to check by giving you some assignment or by giving you case study or giving you a at random problem. Second thing people are trying to assess is that, "Hey, have you done some kind of system design work?" Because still like there is a lot to even understand these things. So, they will ask you questions on, "Hey, design a system for me. Here is a system design where you think the AI should

所以你大概需要九到十周时间，认认真真地把用 AI 做东西、成为一个 builder PM 这条路走一遍。好，假设你已经成了 builder PM，现在你要去面试这个岗位。有了 AI 之后，PM 面试发生了哪些变化？在新的 AI PM 面试里，大家应该预期会遇到什么？是这样，我特别想跟所有人分享一件事，因为我在这方面做了大量研究，每天都有很多人来找我。我的日历被一个个 15 分钟的电话排满了，我会跟我们训练营的成员做这种聊，每次就 15 分钟，因为他们要面试。如果你上过我们的训练营，我是不收费的，因为我就是想帮忙，同时我自己也想跟上现在到底在发生什么。所以让我告诉你正在发生的三件事。第一件，尤其是 level 5、level 6 的 AI 岗位，那种纯考产品直觉的做法已经没了。他们会直接给你一个问题，让你去解，要么是个 case study，要么就在面试现场解。这正在变成一种固定套路。在这个过程里，他们想看的是：'嘿，你到底懂不懂我们现在所处的位置、这个世界是怎么运转的，还是说你还卡在过去——比如半年前、一年前的认知里？你是那个能把我们带向新世界的人，还是会拖着我们用老办法做老决策的人？'这就是我要通过给你布置作业、给你 case study、或者随便扔个问题给你来考察的第一件事。第二件，大家想评估的是：'嘿，你做过系统设计这类活儿吗？'因为光是理解这些东西本身就还有很多门道。所以他们会问你：'嘿，帮我设计一个系统。这是一个系统设计，你觉得 AI 应该在哪里

[1:23:46] Mahesh

improve with open claw or where within Claude code based agentic loop, how will you redesign the system?" Which people sometimes PM just come to me and they are like you know, we have this story where uh my wife was interviewing for an MBA job after doing her MBA. She's a very good software engineer and uh this was like a hybrid job and the and the recruiter asked her, "What are linked list or how to reverse a linked list?" Which is a very basic question for engineers. But she was like expecting an MBA question so she put down the phone. She's like, "I did the whole MBA to get away from linked list and here they are again."

用 OpenClaw 来改进，或者在基于 Claude Code 的 agentic loop 里，你会怎么重新设计这套系统？'有时候 PM 来找我，就会说，你知道吗，我有这么个故事——我太太读完 MBA 之后去面一个 MBA 的岗位。她本身是个很厉害的软件工程师，那是个偏混合性质的岗位，结果招聘的人问她：'什么是链表，怎么反转一个链表？'这对工程师来说是个特别基础的问题。但她当时是在等着回答 MBA 类的问题，所以直接把电话挂了。她说：'我读了整个 MBA 就是为了摆脱链表，结果它们又冒出来了。'

[1:24:25]

[clears throat]

[清嗓子]

[1:24:25] Mahesh

So, some people get offended with these like, "Hey, why is a system design question for me in an AI PM interview or a PM interview?" And becoming normal because if you don't understand the design of these systems, you are going to not find the right capabilities that we should be building on. If you these tools are coming up every day and each design is elevating what's possible. But if you don't understand the design you can't see what the possibilities are and that's why people are trying to test you there. So, these are the two things and beyond that of course like great product sense, great taste in the product, paying attention to detail those are not going anywhere. But these are the two two new things that I will add in testing whether you're a builder PM or not. And I do that all the time like I if I give you a job and if you're not pulling out your Claude code or some kind of a tool like lovable, you're already out. Like if you just ask me like, "Hey, can I do a drawing tool or can I create mocks in Figma?" Those are the things that are like done done. I'm not interested. So, that's the new world. One of the distinctions that you make pretty frequently is this distinction between agentic AI versus AI specifically. So, what is the difference? What do people need to understand? Yeah, so AI is this idea that you can find patterns or this idea that we all have data, machine learning helps you find patterns in data, AI helps you use that day those patterns and make money. Broader AI like AI as a umbrella. And then agentic AI is the thing which allows you to actually take actions, do jobs and finish work. So, the idea is that agentic AI need to have like these three or four components which we talked about. Can I understand the world that I am living right now?

所以有些人会被这种问题冒犯到，心想：'嘿，AI PM 面试、或者 PM 面试里，为什么要给我出系统设计的题？'但这正在变得稀松平常，因为如果你不懂这些系统是怎么设计的，你就找不到我们真正该去构建的那些能力点。这些工具每天都在冒出来，每一次设计上的进步都在抬高'什么是可能的'这条线。但如果你不懂设计，你就看不到有哪些可能性，所以大家就想在这点上考你。这就是两件事，除此之外，当然了，出色的产品直觉、对产品的好品味、对细节的关注，这些都不会消失。但这是我会额外加进去、用来判断你是不是 builder PM 的两件新东西。而且我一直都这么做——如果我给你一个任务，你不掏出你的 Claude Code 或者像 Lovable 这样的工具，你就已经出局了。如果你只会问我：'嘿，我能用个画图工具吗，或者我能在 Figma 里做原型吗？'这些东西就是彻底过去式了，我不感兴趣。这就是新世界。你经常会做的一个区分，是 agentic AI 和单纯的 AI 之间的区别。那这两者的差别是什么？大家需要理解什么？嗯，AI 就是这么个概念：你可以从数据里找出规律。我们都有数据，机器学习帮你在数据里找规律，AI 帮你利用这些规律去赚钱——这是广义的 AI，把 AI 当作一把大伞。然后 agentic AI 是那个真正能让你去采取行动、去干活、去把工作做完的东西。所以核心是，agentic AI 需要具备我们刚才聊到的这三四个组成部分。我能不能理解我现在所处的这个世界？

[1:26:23] Mahesh

Can I understand what's happening right now in that world? Building my context with these two. Once I have built can I take actions? Which is back running bash commands or calling tools or MCP servers. When I do that can I run my own evals and make sure that I have achieved the goal or not? And if I have done all these three then I'm an agentic AI or product. Or if I just send you something like, "Hey, is it positive negative emotion?" then I'm doing more like an AI thing which is I can do one thing specifically and if you send me in this format then I will work. Else best of luck. That's the whole world of AI or cognitive services we used to talk in Microsoft about. But this is like a world where we are relying a lot on the model or this intelligence and giving it a loosely connected tools, knowledge and memory and then just trusting it to solve world hunger or any problem thrown at it. That's the agentic AI. And that's where most of the excitement and money is today. I want to ask you a couple hot personal questions. You spent 13 years in big tech. You started in Microsoft around 2012. You left to Google in 2025. Can you share the honest what can people accept expect in terms of total compensation trajectory? What was yours over those 13 years? Yeah, I think first it's pretty standard right to use start with 120. I think AI worked very well for me. So, I started with 120, spent all my life at Microsoft to grow it at 360, 400 and that time I felt like I have achieved nirvana. This is like the best it can be and then when uh I think it got a 70% bumped when I joined Meta. And then another 70%. So, this is all the AI that was there. So, after that I pretty much doubled my salary every year. Every two years, 18 months, every switch I made was a double salary switch.

我能不能理解此刻这个世界里正在发生什么？用这两点来构建我的 context。一旦构建好了，我能不能采取行动？也就是去跑 bash 命令、调用工具、或者调用 MCP server。做完这些之后，我能不能跑自己的 evals，确认我到底有没有达成目标？如果这三件事我都做到了，那我就是一个 agentic AI 或者 agentic 产品。反过来，如果你只是给我发点东西，比如'嘿，这是正面还是负面情绪？'那我做的就更像是单纯的 AI——我只能专门做一件事，而且你得按这个格式发给我我才能干活，否则就自求多福吧。这就是过去我们在微软常说的那个 AI、或者说认知服务的世界。但现在这个世界，我们非常依赖模型、依赖这种智能，给它一套松散连接的工具、知识和记忆，然后就这么信任它去解决世界饥饿问题、或者任何抛给它的难题。这就是 agentic AI，也是当下大部分兴奋点和金钱所在的地方。我想问你几个比较辛辣的个人问题。你在大厂待了 13 年，2012 年左右从微软起步，2025 年离开去了谷歌。能不能坦白讲讲，大家在总薪酬走势上可以预期什么？你这 13 年自己是怎么走过来的？嗯，我觉得首先，起步薪酬挺标准的，对吧，从 12 万美金开始。我觉得 AI 对我帮助特别大。所以我从 12 万起步，在微软待了大半辈子把它涨到 36 万、40 万，那个时候我觉得自己已经修成正果了，觉得就这样了、到顶了。然后我跳去 Meta 的时候涨了大概 70%，接着又涨了 70%。这全是当时那波 AI 带来的。所以那之后我基本上每年工资翻一倍——每两年、18 个月，我每跳一次槽就是翻倍。

[1:28:22] Mahesh

So, if you and my last comp it was looking at 1.3, 1.4 million is what you make easy at my level in my experience if you're working in AI. And this is not you applying for jobs, this is them saying, "Hey, we need you. We are doing this new thing. Seems like you are the only one who have done this before. You tell us what you're getting. We will give you 30 40% on top of whatever you're making." And then you can say I need 100% and then generally they don't say no. So, and this is not only my story right all my friends who were at any stage right if they were in Meta, they are working in Nvidia today and their total comp is looking at two 2.5 million. Wow, that's insane. So, is that why you bounced around so much cuz I never see people who worked at all four companies.

所以呢，如果你……我最后一份薪酬大概在 130 万、140 万美金，以我这个级别和经验，在 AI 领域工作的话这是很轻松能拿到的数。而且这还不是你去投简历求职，而是他们主动说：'嘿，我们需要你。我们在做一件新东西，看起来你是唯一一个以前干过这事的人。你说你现在拿多少，我们在你现有的基础上再给你加 30%、40%。'然后你可以说我要加 100%，一般他们也不会拒绝。所以这不光是我一个人的故事，对吧——我所有朋友，不管处在哪个阶段，如果他们当时在 Meta，今天就在 Nvidia 干活，总薪酬大概在 200 万、250 万美金。哇，这太疯狂了。所以你是因为这个才跳来跳去的吗？因为我从没见过有人四家公司全待过。

[1:29:12]

[laughter]

[笑声]

[1:29:12] Mahesh

I wish but it's not that right there I like I I loved Meta. So, Meta was because I was just bored at Microsoft and I wanted to get out of Seattle because of personal reasons. I wanted to live in Bay Area. And so, Meta was a great company and I would never left Meta, but Meta had this legal visa problem. So, once I switch, I needed like my green card and it need to switch my roles as like a product manager. And they couldn't do it because they had a USCIS case pending. So, they couldn't file like my green card and I was in that line for 10 years. You might have known this little that we stay. That's like one life thread that always running. So, I gave them like good 8 months to a year to resolve that USCIS case, but it was not moving anywhere. At that time, I needed to make a call. At the same time, I could admit that Bedrock at AWS was just building and they needed somebody and they threw a lot of money at me. So, I could have waited more, but they didn't let me wait at all. So, that's was big deal. And then, Google was more like a dream company, to be honest. Like my wife always told me that if you are a PM and if you have not worked at Google, you are not a PM. Like that was her way of judging me. Because, you know, if you you've done an MBA, so you need not to prove to the world. When you come from an engineering become background and you become a product manager, you have to prove to the world that you are a legit product manager. You're not a developer just wore a suit. So, that was just that and when I they needed somebody who has built frameworks and actually build agents in 2024. And that I think what got me and not my like awesome frameworks or Porter five forces. It was mostly how much I knew about AI and so was my interview. My

我倒希望是这样，但其实不是。我是真的很喜欢 Meta。去 Meta 是因为我在微软待得太无聊了，而且出于个人原因想离开西雅图，我想住在湾区。Meta 是家很棒的公司，我本来永远不会离开 Meta，但 Meta 当时有个法律上的签证问题。我一旦跳槽，就需要办我的绿卡，还得把我的角色切换成产品经理，而他们办不了，因为他们有一个 USCIS 的案子还在挂着。所以他们没法帮我递交绿卡申请，而我已经在那个队里排了 10 年了。你可能不太了解这事，但这就像一条贯穿一生、始终悬着的线。所以我给了他们足足八个月到一年时间去解决那个 USCIS 的案子，但它压根没往前推进。那个时候我得做个决定。与此同时，我能看到 AWS 的 Bedrock 当时正在搭建，他们需要人，而且砸了一大笔钱给我。我本来可以再等等，但他们根本不让我等。所以这是个大事。然后呢，说实话，谷歌更像是个梦想中的公司。我太太一直跟我说，如果你是个 PM 却没在谷歌待过，那你就不算 PM。这就是她评判我的方式。因为你想，如果你读过 MBA，你就不用再向世界证明什么。但当你是工程背景出身、转做产品经理，你就得向世界证明你是个正经的产品经理，而不是个穿了西装的开发。就是这么回事。当时他们需要一个搭过框架、并且在 2024 年真正构建过 agent 的人。我觉得打动他们的就是这点，而不是我那些所谓厉害的框架或者波特五力。主要是我对 AI 懂得有多深，我的面试也是这样。我的

[1:31:00] Mahesh

interview was very AI driven, like very what I have done in AI and how I can bring AI to production. So, that was my Google. So, first that comp and I think it was never driven by money. Otherwise, I would have never left, right? So, Why did you leave 1.3 at Google when you left and it would be something like 2.5, 2.6 now if you stayed and maybe jumped again? So, why did you leave? What are you up to now? Yeah, so idea is like I think these companies are going to throw a lot of money at you to keep you and then waste you. So, that's was my observations, especially at Google, right? With me. I think I'm away from them enough and I'm not going back ever. So, the idea is that these are large companies. And if you look at like what happened in AI is large companies have not produced in AI. Like if you look at OpenAI, it was a small company which created ChatGPT. Then it was a small company Lovable that created Lovable. It didn't came out of Google. And then, if you look at Claude code, it was created by a very small team inside Anthropic when Anthropic was not big. And then open claw. So, what has happened in with AI is the tools are distributed, but these big companies have no environment to grow something which is can be imagined, put in production and put in customer hands. I'm pretty sure. I love people at Google. Most of the level three thinkers or level four thinkers live inside Google. I will go any day to stay even two hours with those people. I love them. But, the company will never launch something like open claw. This something like this will be killed. Maybe you are thrown out of the company for trying something like open claw. So, that's the kind of environment and that's the kind of guardrails they have put inside their whole ecosystem because

面试是非常 AI 导向的，全在问我在 AI 上做过什么、我怎么把 AI 带到生产环境。这就是我进谷歌的过程。所以首先是那份薪酬，而且我觉得这从来都不是钱驱动的，否则我根本不会离开，对吧？那你为什么要走呢？你离开谷歌的时候是 130 万，如果你留下来、说不定再跳一次，现在大概能到 250 万、260 万。所以你为什么要走？你现在在忙什么？嗯，我的想法是这样的——我觉得这些公司会砸一大笔钱把你留住，然后把你浪费掉。这是我的观察，尤其是在谷歌，对我来说就是这样。我觉得我离他们已经足够远了，也永远不会回去了。核心在于，这些是大公司。你看看 AI 领域发生了什么——大公司在 AI 上没什么产出。比如 OpenAI，它是一家小公司，做出了 ChatGPT。然后是 Lovable 这家小公司，做出了 Lovable，它不是从谷歌出来的。再看 Claude Code，它是 Anthropic 内部一个非常小的团队做出来的，那会儿 Anthropic 还不大。然后还有 OpenClaw。所以 AI 领域发生的事情是，工具是分散的，而这些大公司根本没有那种环境，去培育出一个能被想象、能投入生产、能真正交到客户手里的东西。我很确定这点。我爱谷歌的那些人，大多数 level 3、level 4 的思考者都在谷歌内部。我随时都愿意，哪怕只是为了跟那些人待上两小时也行，我爱他们。但这家公司永远不会推出像 OpenClaw 这样的东西，这样的东西在里面会被扼杀掉。你甚至可能会因为尝试做像 OpenClaw 这样的东西而被赶出公司。所以这就是那种环境，那种他们在整个生态系统里设下的护栏，因为

[1:32:58] Mahesh

it's such a big machine. Same as ChatGPT, right? Like if you look at OpenAI today, they have become so big and it's very hard for you to just come up with new ideas and throw it on Twitter and then take feedback and iterate for 6 months and then one day say I have created something which is like some I think largest liked repo on GitHub. That is not possible and for me, I reached a point in my life that I wanted to just stay unbounded as much as I could and I was very blessed that I had this course and that was giving me lot of satisfaction of staying with people and not losing like one thread I had like money was like secondary, to be honest. Main problem was that you you get so dependent on these institutions for learning, for staying up to date. I do miss, right? I will pay them today to be around Google, right? So, that is the main thing, but for me, that course that I started teaching on Maven always helped me stay even up to date than what is there. So, as you see, I have tried open claw. I have built my mini PM. I would have done more here than what I would have done at Google. And given a choice, I will never go go to Google again or to be honest, any company. It just they just kill every intelligence neuron you have with Like you won't believe that, you know, for a two-page document, you have to have a one page of approvals. And that takes like 6 weeks. In 6 weeks, a non-builder PM becomes a builder PM. And then they can build anything they want ever. So, that's the world we live in and these are the companies

它就是台这么大的机器。ChatGPT 也一样，对吧？你看今天的 OpenAI，他们已经变得太大了，你很难就这么冒出个新点子、扔到 Twitter 上、收集反馈、迭代六个月，然后某一天说我做出了一个东西——我想大概是 GitHub 上获赞最多的 repo 之一。这是不可能的。而对我来说，我到了人生中这么一个节点，我就想尽可能地保持不受束缚的状态。我很幸运手里有这门课，它给了我很大的满足感——能跟人待在一起，又不丢掉我那一根重要的线。说实话，钱对我来说是次要的。主要的问题是，你会变得太依赖这些机构来学习、来保持与时俱进。我确实会怀念，对吧？我今天甚至愿意花钱去待在谷歌身边。所以这是主要的事。但对我来说，我在 Maven 上开始教的那门课，一直帮我保持着比待在那里还要更与时俱进的状态。所以你看，我试过 OpenClaw，我搭了我自己的 mini PM。在这里我做成的事，比我留在谷歌能做的还要多。如果让我选，我永远不会再回谷歌，说实话，也不会再回任何公司。它们就是会把你每一个智慧的神经元都扼杀掉。你可能不敢相信，你知道吗，为了一份两页的文档，你得拿到一页的审批，而那要花六周。六周时间，一个非 builder PM 都能变成 builder PM 了，然后他就能造出任何他想要的东西。所以这就是我们身处的这个世界，这些就是那些公司