【AI Engineer】Graph Engineering
字幕摘录
| 时间 | 英文 | 中文 |
|---|---|---|
| 0:00 | I came on here to talk about a term I keep seeing going viral on Twitter, it's graph | 我来这是为了谈一个我一直在推特上看到病毒传播的术语 这是图表 |
| 0:05 | engineering. | 工程学。 |
| 0:06 | You've seen it, I've seen it too. | 你也看过,我也看过 |
| 0:08 | And I'll be honest, the first time I saw it, | 老实说 我第一次见到它 |
| 0:11 | my reaction was, okay, is this a real thing? | 我的反应是,好吧,这是真的吗? |
| 0:13 | Or did we just invent another phrase to make everyone | 或者我们只是发明了另一个短语 让每个人 |
| 0:17 | feel behind? | 有感觉吗? |
| 0:18 | Because AI has this funny habit where every few weeks, there's this new term that goes | 因为AI有这种有趣的习惯 每隔几周就有一个新的术语 |
| 0:24 | 病毒(英语:Viral). | |
| 0:25 | Prop engineering, context engineering, agent engineering, vibe coding, loop engineering, | 道具工程 上下文工程 代理工程 大气编码 循环工程 |
展开字幕全文（500 条）
| 序号 | 英文 | 中文 |
|---|---|---|
| 1 | I came on here to talk about a term I keep seeing going viral on Twitter, it's graph | 我来这是为了谈一个我一直在推特上看到病毒传播的术语 这是图表 |
| 2 | engineering. | 工程学。 |
| 3 | You've seen it, I've seen it too. | 你也看过,我也看过 |
| 4 | And I'll be honest, the first time I saw it, | 老实说 我第一次见到它 |
| 5 | my reaction was, okay, is this a real thing? | 我的反应是,好吧,这是真的吗? |
| 6 | Or did we just invent another phrase to make everyone | 或者我们只是发明了另一个短语 让每个人 |
| 7 | feel behind? | 有感觉吗? |
| 8 | Because AI has this funny habit where every few weeks, there's this new term that goes | 因为AI有这种有趣的习惯 每隔几周就有一个新的术语 |
| 9 | 病毒(英语:Viral). | |
| 10 | Prop engineering, context engineering, agent engineering, vibe coding, loop engineering, | 道具工程 上下文工程 代理工程 大气编码 循环工程 |
| 11 | and now graph engineering. | 而现在的图表工程。 |
| 12 | Some of these phrases are hype. | 其中一些短语是杂音. |
| 13 | Some of them are actually useful. | 其中一些实际上是有用的。 |
| 14 | And | 还有 |
| 15 | graph engineering is one of the useful ones, because it gives you a much better way to think | 图表工程是有用的之一,因为它给你更好的思考方法 |
| 16 | about how AI actually gets done. | 关于AI如何真正完成。 |
| 17 | So in this episode, I'm going to explain graph engineering | 所以在这集里,我要解释一下图表工程 |
| 18 | in plain English. | 简单英语。 |
| 19 | By the end of this episode, I want you to be able to take one AI workflow | 到本集结束,我要你能够 采取一个AI工作流程 |
| 20 | you already run, like customer research, support triage, content production, or startup idea | 您已经运行, 如客户研究, 支持分类、 内容制作或启动想法 |
| 21 | validation, and turn it into a simple map of steps, checks, handoffs, loops, and human approvals. | 验证,并变成简单的步骤图,检查图,交割图,循环图,以及人类的审批图. |
| 22 | So we're going to talk about all that and how you can do it. | 所以,我们要谈论这一切 和你如何做到这一点。 |
| 23 | It's going to be clearly explained. | 将作明释. |
| 24 | So let's get into it. | 因此,让我们开始吧。 |
| 25 | The simplest way to think about graph engineering is like this. | 最简单的思考图形工程的方法就是这样. |
| 26 | Prompt engineering is how you ask the AI for a better question. | 即时工程就是问AI一个更好的问题. |
| 27 | And context engineering is | 环境工程 |
| 28 | how you give AI better information. | 你如何给AI更好的信息。 |
| 29 | But graph engineering is how you design the work around | 但图表工程是设计工作的方式 |
| 30 | the AI so the whole thing stops living inside one messy, giant AI chat. | 所以整个事情都不再 生活在一个混乱,巨大的AI聊天。 |
| 31 | I'll give you an example. | 举个例子 |
| 32 | Imagine you're researching a new startup idea. | 想象一下你在研究一个新的启动想法。 |
| 33 | The normal way most people use AI is they open | 大多数人使用AI的正常方式是打开 |
| 34 | up a chat and they say, should I build this idea? | 闲聊一下,他们说, 我应该建立这个想法吗? |
| 35 | The model will give you a confident answer. | 模型会给你一个自信的答案。 |
| 36 | It probably sounds pretty smart. | 听起来很聪明 |
| 37 | It might give you the market size, a few competitors, maybe a | 它可能会给你市场规模, 一些竞争者,也许一个 |
| 38 | market plan, and you feel like you did the research. | 市场计划,你觉得你做了研究。 |
| 39 | But if you actually slow down, | 但是如果你真的慢下来 |
| 40 | you realize something a little uncomfortable happened. | 你知道有些事情有点不舒服 |
| 41 | One model in one past decided what | 过去一个模型决定 |
| 42 | mattered, researched the market, interpreted the evidence, wrote the recommendation, | 研究市场 解释证据 写建议 |
| 43 | and graded it in its own confidence. | 并秘密地加以分级。 |
| 44 | That's a lot of trust to put into one blob of text. | 满嘴的文字都是信任 |
| 45 | In some cases, you might spend years of your life based on this one question that you asked, | 在某些情况下,你可能会度过你生命中的岁月 基于这个你问的问题, |
| 46 | and you might be working on the wrong thing. | 你可能搞错了 |
| 47 | The graph version looks a lot different. | 图表版本看起来大不相同. |
| 48 | So a planner first breaks the question into angles. | 因此,一个计划者首先把问题分成了角度. |
| 49 | One researcher looks at the customer, | 一个研究员看着顾客 |
| 50 | another looks at competitors, another looks at distribution, another looks at pricing, | 另一种看竞争对手,另一种看分配,另一种看定价, |
| 51 | another looks at risks. | 又一看风险。 |
| 52 | Then a skeptic will try to kill the weak findings. | 然后怀疑者会试图杀死那些薄弱的发现. |
| 53 | Then a merger turns | 然后合并又开始 |
| 54 | their surviving evidence into a one page recommendation. | 他们幸存的证据 在一个页的建议。 |
| 55 | And then you approve the | 然后,你批准 |
| 56 | decision before you act on it. | 在你采取行动之前作出决定。 |
| 57 | The output might still be this written report, but the work behind | 产出可能仍然是这份书面报告,但背后的工作 |
| 58 | it is just designed so much better. | 它的设计好得多。 |
| 59 | And that at its core is graph engineering. | 其核心是图表工程。 |
| 60 | You're taking a | 你拿一个 |
| 61 | task and turning it into a workflow that you can actually manage. | 任务,并把它变成一个可以实际管理的工作流程。 |
| 62 | Now let's define the basic | 现在,让我们定义基本 |
| 63 | vocabulary without making this feel like a computer science lecture. | 词汇并不让人觉得这像一个计算机科学讲座。 |
| 64 | By the way, I remember learning | 顺便说一句,我记得学习 |
| 65 | about one of my first classes in university was graph theory. | 我大学的一门课是图论 |
| 66 | And so it's a real throwback for me. | 所以对我来说是真正的回击 |
| 67 | I'll explain it to you in the clearest way possible. | 我会用最清楚的方式解释给你听的 |
| 68 | When people say graph, they basically | 当人们说图表,他们基本上 |
| 69 | mean jobs connected by arrows. | 由箭头连接的工作。 |
| 70 | Each job is a step in the workflow. | 每项工作都是工作流程中的一个步骤. |
| 71 | The arrows show | 箭头显示 |
| 72 | what happens next. | 接下来会发生什么? |
| 73 | And the shared notes moving through the workflow are the state, | 共享的音符通过工作流程移动是状态, |
| 74 | which is just a fancy way of saying, what does the system know so far? | 这仅仅是一种美妙的说法, 系统知道什么到目前为止? |
| 75 | So that sounds technical | 听起来很技术 |
| 76 | for about five seconds. | 大约5秒钟。 |
| 77 | And then you realize that's actually how work gets done in the real world. | 然后你意识到 现实世界的工作就是这样完成的。 |
| 78 | In reality, think about customer support. | 现实中,考虑客户支持. |
| 79 | When a customer writes in, the work is rarely just | 当顾客写到,作品很少只是 |
| 80 | answer the ticket. | 接电话 |
| 81 | First, you need to understand what kind of issue it is. | 首先,你需要了解它是什么问题。 |
| 82 | Then you need to | 那你必须 |
| 83 | check the customer's account history. | 检查客户的账户历史。 |
| 84 | Maybe you need to search for the docs for the right policy. | 也许你需要寻找文件 正确的政策。 |
| 85 | Then you draft a response. | 然后你起草一个答复。 |
| 86 | Then you decide whether this is risky enough that a human should review | 那你决定这是否够冒险 让人类去审查 |
| 87 | it before going out. | 在出门前 |
| 88 | When you draw those steps out and connect them in an order, | 当你画出这些台阶 并连接它们在一个顺序, |
| 89 | they actually depend on each other. | 他们实际上是互相依赖的。 |
| 90 | And that is a graph. | 这是一个图表。 |
| 91 | Take content, for example. | 以内容为例. |
| 92 | If I'm | 狦и琌 |
| 93 | making a YouTube episode, the work isn't just write a script. | 制作一个YouTube插曲, 作品不只是写一个剧本。 |
| 94 | A good episode might start with | 一个好的剧集可能从 |
| 95 | research, a thesis, examples, a hook, maybe a script, then title ideas, then | 研究 论文 实例 钩子 可能还有剧本 然后是标题 |
| 96 | thumbnail directions, then an Excalibur draw, and then a final pass where I ask, | 缩略图方向,然后是Excalibur图,然后是最后的通行证,我问, |
| 97 | does this sound like a human being or does this sound like someone trapped inside a SAS onboarding | 这听起来像一个人 或者这听起来像有人 困在SAS的船上 |
| 98 | flow? | 流动? |
| 99 | Some of those steps have to happen in order. | 有些步骤必须有序进行。 |
| 100 | You probably want the thesis before the script. | 你可能想要论文 在剧本之前。 |
| 101 | You probably want the script before the Excalibur draw. | 你可能想要剧本 在Excalibur画前。 |
| 102 | But other pieces can happen at the same | 但是,其他的片段 可能发生在同一时间 |
| 103 | time. | 时间。 |
| 104 | One researcher can look for examples while another looks for counterarguments. | 一名研究人员可以寻找实例,另一名研究人员则寻找反证。 |
| 105 | One could study the audience angle while another looks for practical workflows. | 人们可以研究观众角度,而另一个则寻找实际的工作流程。 |
| 106 | Then those outputs | 然后这些产出 |
| 107 | merge back into the script. | 合并到脚本中。 |
| 108 | And that's where the graph starts paying. | 这就是图开始支付。 |
| 109 | Because most people use | 因为大多数人使用 |
| 110 | AI in a straight line because chat makes everything kind of feel sequential. | 直线的AI,因为聊天让一切感觉都是顺序的. |
| 111 | You ask for research, | 你问研究, |
| 112 | then you ask for summary, then you ask for a draft, and then you ask for edits, then you ask for | 然后要求摘要,然后要求草稿,然后要求编辑,然后要求编辑 |
| 113 | titles. | 标题。 |
| 114 | That works for really simple things. | 这对很简单的事情是有用的。 |
| 115 | But when the work has multiple pieces, the straight | 但是,当工作有多个片段,直 |
| 116 | line chat starts to get slow and fuzzy and actually hard to trust. | 线条聊天开始变得缓慢和模糊,实际上难以信任。 |
| 117 | What's cool about a graph is it | 一个图最酷的是 |
| 118 | lets you design the work more like a small team. | 让你设计的工作更像一个小团队。 |
| 119 | One part plans, a few work in parallel, another | 一部分计划,几项平行工作,另一部分 |
| 120 | checks the work, another merges it, and then the human approves the final step. | 检查工作,另一个合并, 然后人类批准最后一步。 |
| 121 | And once that | 一旦这样 |
| 122 | clicks in your head, it just gets a lot less mysterious. | 点击你的头, 它只是得到很多的神秘。 |
| 123 | Because there's two different things | 因为有两件事不同 |
| 124 | people mean when they say graph in AI. | 人们在AI里说图表时的意思 |
| 125 | And this is actually where a lot of the confusion comes from. | 这其实就是很多混乱的根源所在。 |
| 126 | The first is what's called a knowledge graph. | 第一是所谓知识图. |
| 127 | A knowledge graph helps AI reason over relationships | 一个知识图可以帮助AI理性战胜关系 |
| 128 | over things. | 对事物。 |
| 129 | For example, this customer works at this company, this company uses this product, | 例如,这个客户在这个公司工作, 这家公司使用这个产品, |
| 130 | this product connects to this tool, this support issue relates to this feature, and this feature | 此产品连接到此工具, 此支持问题与此特性有关, 而此特性 |
| 131 | is owned by this team. | 属于这个团队。 |
| 132 | Knowledge graphs help because AI reason across relationships and messy | 知识图有帮助 因为人工智能的理性 跨越关系和混乱 |
| 133 | data. | 数据。 |
| 134 | This matters because normal rag often retrieves chunks of text that look similar to | 这很重要,因为普通的布往往会检索到看起来类似于 |
| 135 | question, but it can struggle when the answer actually requires connecting different people | 问题,但它会挣扎 当答案实际上需要 连接不同的人 |
| 136 | across companies and topics and claims and events. | 跨公司、专题、权利主张和活动。 |
| 137 | You know, there's tools like you might have heard | 你知道,有 工具像你可能已经听说过 |
| 138 | of Microsoft graph rag, because sometimes you just need AI to understand relationships inside | 微软图布,因为有时你只需要AI来理解内部的关系 |
| 139 | a body of knowledge, not just to retrieve the nearest paragraph. | 一个知识体, 不只是检索最近的段落。 |
| 140 | That is one version of graph | 这是图的一个版本 |
| 141 | engineering. | 工程学。 |
| 142 | The second version is what's called an agent graph. | 第二个版本是所谓的代理图. |
| 143 | An agent graph is about how work | 一个代理图 是关于如何工作 |
| 144 | moves. | 移动。 |
| 145 | So a planner hands work to researchers, the researchers work in parallel, a skeptic | 所以计划师的手对研究者起作用 研究者平行工作 一个怀疑者 |
| 146 | checks the findings, a synthesizer might merge the parts, and a human will, you know, prove the | 检查发现,一个合成器 可能合并部分, 和人类的意愿,你知道,证明 |
| 147 | final answer. | 最终答案。 |
| 148 | This episode is mostly about agent graphs actually, because that is the version you | 这集主要是关于代理图表 因为这就是你的版本 |
| 149 | start using today as a founder, as a creator, as an operator, as a small team, so I figured I'd do | 开始使用今天作为一个创始人,作为一个创造者,作为一个操作员,作为一个小团队, 所以我想我会做 |
| 150 | an episode focusing on that. | 一集的焦点。 |
| 151 | The easiest way to remember the difference, though, is kind of like | 记得区别的最简单的方法是 |
| 152 | this. | 这个 |
| 153 | Knowledge graphs help AI understand how information connects, whereas agent graphs help AI | 知识图帮助AI理解信息如何连接,而代理图帮助AI |
| 154 | understand how work should move. | 理解工作应该如何进行。 |
| 155 | And eventually, the truth is the best systems use both. | 而最终,真理是两种方法都使用的最佳系统. |
| 156 | The AI | 大赦国际 |
| 157 | will understand relationships inside your business, and it will also know how to move through the | 将了解你业务中的关系,它也会知道如何通过 |
| 158 | right steps. | 正确的步骤。 |
| 159 | But how can we make this tactical? | 但是,我们怎么才能使这个战术? |
| 160 | When should you use graph engineering? | 你什么时候用图表工程? |
| 161 | Well, use | 来,用 |
| 162 | it when the work has multiple steps, multiple sources, maybe multiple paths, checks, risk or | 当工作有多个步骤,多个来源,也许有多个路径,检查,风险或 |
| 163 | approvals. | 核准。 |
| 164 | Honestly, if you're asking AI to brainstorm 10 names for a new project, you probably | 老实说,如果你要求AI 集思广益10个名字 一个新的项目,你可能 |
| 165 | don't need a graph. | 不需要图表 |
| 166 | If you're asking AI to summarize a short email, you probably don't need a graph. | 如果你要求AI总结一个简短的电子邮件,你可能不需要图表. |
| 167 | But if you're using AI to do deep research, create a go-to-market plan, triage support tickets, | 但如果你用AI做深入研究 创造出上市计划 分批支持票 |
| 168 | review code, prepare for sales calls, synthesize customer feedback, or produce recurring content | 审查代码、准备销售电话、综合客户反馈或制作经常性内容 |
| 169 | workflow, that's when graph thinking actually starts to matter a lot. | 工作流程,也就是当图形思维开始变得重要的时候。 |
| 170 | And the rule is pretty | 规则是漂亮的 |
| 171 | simple. | 很简单 |
| 172 | Use a graph when the work has multiple steps, some steps can happen at the same time, | 当工作有多个步骤时使用一个图表,一些步骤可以同时发生, |
| 173 | and the final output needs checking before it matters. | 最终产出需要先检查 才会重要 |
| 174 | A diamond starts with one question, | 钻石从一个问题开始 |
| 175 | splits into multiple parallel paths, checks the work, and then merges everything back into one | 分割成多个平行路径, 检查工作, 然后将所有内容合并到一个 |
| 176 | answer. | 答问. |
| 177 | So here's the startup idea version. | 这就是启动想法的版本。 |
| 178 | Let's say the question is, should I launch an AI | 假设问题是,我是否应该启动一个AI |
| 179 | bookkeeping product for Shopify merchants? | 商店商人的记账产品? |
| 180 | The messy chat version is one big question | 乱谈版是个大问题 |
| 181 | and one big answer. | 和一个大答案。 |
| 182 | The graph version starts with a planner. | 图文版本从计划员开始. |
| 183 | So the planner would say something like, | 所以策划者会说一些类似的话, |
| 184 | to answer this well, we need to understand the customer pain, the competitive landscape, | 要很好地回答这个问题,我们需要理解 顾客的痛苦,竞争环境, |
| 185 | the go-to-market wedge, the pricing pressure, and the risks. | 上市的楔子、定价压力和风险 |
| 186 | And then the work splits. | 然后工作分开。 |
| 187 | You have one | 你有一个 |
| 188 | researcher who studies Shopify merchants and tries to understand the bookkeeping pain. | 研究 Shopify商家的研究员 试图理解书记的痛苦 |
| 189 | Are | 对 |
| 190 | they using QuickBooks? | 他们用快书? |
| 191 | Are they using spreadsheets? | 他们在用电子表格吗? |
| 192 | Are they hiring bookkeepers? | 他们在雇书记员吗? |
| 193 | Are they annoyed at | 他们生气吗? |
| 194 | tax time? | 课税时间? |
| 195 | Are they looking for automation? | 他们在找自动化吗? |
| 196 | Or do they just want someone to clean up the mess | 还是他们只是想找人清理烂摊子 |
| 197 | once a month? | 一个月一次? |
| 198 | You have another researcher who's stunning competitors. | 你还有一个研究者 谁惊人的竞争对手。 |
| 199 | 是否已经存在 shostify | |
| 200 | bookkeeping tools? | 记账工具? |
| 201 | Are there accounting firms building this manually? | 是否有会计师事务所手工建造? |
| 202 | 是 App Store 产品 | |
| 203 | solving this at all? | 完全解决了吗? |
| 204 | Are freelancers on Upwork or Fiverr doing the work in a way that software | 自由职业者在Upwork或Fiverr 工作的方式 软件 |
| 205 | could partially replace? | 可以部分替换吗? |
| 206 | Maybe you have another researcher who's studying the distribution. | 或许你还有一位研究发行的研究员 |
| 207 | Shopify商人们在哪里混? | |
| 208 | What newsletters do they read? | 他们读什么通讯? |
| 209 | What agencies already | 哪些机构已经 |
| 210 | have trust with them? | 信任他们吗? |
| 211 | What Shopify app categories do they search? | 他们搜索什么软件分类? |
| 212 | What search terms reveal buying | 搜索条件显示购买 |
| 213 | intent? | 意图? |
| 214 | Those three jobs can happen at the same time because they don't depend on each other. | 这三项工作可以同时发生,因为它们不依靠对方. |
| 215 | Then comes the skeptic. | 然则疑者来也. |
| 216 | The skeptic asks, what claims are actually supported? | 怀疑论者问,什么主张得到实际支持? |
| 217 | Which evidence | 哪些证据 |
| 218 | is stale because you're going to have data that is just old? | 因为数据太老化了,所以太僵化了? |
| 219 | Which competitor is being ignored? | 哪个竞争者被忽略了? |
| 220 | Where are we confusing pain with willingness to pay? | 我们在哪里把痛苦和愿意付出混淆? |
| 221 | Where did the AI sound confident without | AI从哪里听来没有自信 |
| 222 | proving anything? | 证明什么? |
| 223 | And this step matters more than people think. | 这个步骤比人们想象的更重要 |
| 224 | A lot of AI research fails | 许多AI研究失败 |
| 225 | because the same model that writes the answer also grades the answer. | 因为写答案的模型 也给答案分级 |
| 226 | That is like asking someone | 这就像问别人 |
| 227 | to write their own performance review and then being shocked when they describe themselves | 写自己的考绩报告,然后在描述自己时感到震惊 |
| 228 | as a visionary. | 作为一个有远见的人。 |
| 229 | Come on. | 来吧。 |
| 230 | In a good graph, checking is its own job. | 在好的图表中,检查是其自身的工作. |
| 231 | Then comes the merge. | 然后是合并。 |
| 232 | The merge step takes the surviving evidence and turns it into a recommendation. | 合并步骤将幸存的证据变成建议. |
| 233 | Should we pursue | 我们是否应该追求 |
| 234 | this? | 这个? |
| 235 | Should we pause it? | 我们要暂停吗? |
| 236 | Should we kill it? | 我们要杀了它吗? |
| 237 | What is the wedge? | 什么是楔形? |
| 238 | Who's the first customer? | 谁是第一个客户? |
| 239 | What should we test this week? | 我们这周该测试什么? |
| 240 | And what evidence would actually change our mind? | 什么样的证据会改变我们的想法? |
| 241 | And finally, you have the human gate. | 最后,你有人类的大门。 |
| 242 | That's where you decide what to do next. | 这就是你决定下一步要做什么的地方。 |
| 243 | You might decide to | 你可能会决定 |
| 244 | record a landing page teardown of a Shopify merchants. | 记录了一个商店商的登陆页面被撕毁 |
| 245 | You might decide to interview 10 | 你可以决定采访10个 |
| 246 | Shopify agency owners. | 收购代理店主。 |
| 247 | You might decide to build a tiny calculator that estimates bookkeeping | 你可能会决定建立一个小计算器 来估计簿记 |
| 248 | cleanup costs. | 清理费用。 |
| 249 | Or hey, you might decide the idea is way too crowded and you just want to move on. | 或者,嘿,你可能会决定 这个想法太拥挤 你只是想继续前进。 |
| 250 | But that is the point. | 但这就是重点。 |
| 251 | Graph engineering does not magically make the decision for you. | 图形工程不会神奇地为你做决定 |
| 252 | It gives you | 它给你 |
| 253 | a better way to produce the evidence you use to make the decision. | 一个更好的方法 提供你用来做决定的证据。 |
| 254 | Now, this is where I think | 现在,这就是我想 |
| 255 | people get too fancy too quickly. | 人们变得太花哨太快。 |
| 256 | I would start way simpler than you see on Twitter people using | 我开始会比你在推特上看到的 使用 |
| 257 | land graph. | 陆地图. |
| 258 | You see people using autogen or some custom agent framework on day one. | 你看到人们在第一天使用自动或一些自定义代理框架. |
| 259 | For your first | 为了你的第一个 |
| 260 | graph, you can actually run it manually behind the scenes. | 图表,你可以在幕后手动操作。 |
| 261 | I don't know why more people don't do this. | 我不知道为什么更多的人不这样做。 |
| 262 | I could show you exactly how to do it, but that just might be boring. | 我可以告诉你怎么做 但那可能很无聊 |
| 263 | The important thing is the | 重要的是 |
| 264 | structure. | 结构。 |
| 265 | Give each job its own lane. | 给每个工作自己的车道。 |
| 266 | One lane does customer research. | 一条路是客户研究的 |
| 267 | Another lane does competitor | 另一条路是竞争对手 |
| 268 | research. | 调查。 |
| 269 | Another lane does distribution research. | 另一条航道进行分配研究. |
| 270 | Then the checker lane attacks the evidence. | 然后检查车道攻击证据。 |
| 271 | Then | 礛 |
| 272 | merge lane turns the surviving evidence into a recommendation. | 合并道将幸存的证据变成推荐书. |
| 273 | That is already graph engineering. | 这已经是图表工程了。 |
| 274 | It's like level one of graph engineering. | 这就像图工程的一级。 |
| 275 | Yes, it's slower than a fully automated system, | 是的,它比完全自动化的系统慢, |
| 276 | but it's way easier to understand. | 但这样更容易理解 |
| 277 | And if the manual version doesn't produce way better work, | 如果手动版本不能产生更好的效果 |
| 278 | automating it honestly will just produce mediocre work way faster. | 诚实地实现自动化只会更快地产生平庸的工作. |
| 279 | The first rep is to draw | 第一个代表是画画 |
| 280 | the graph before you automate the graph. | 将图表自动化。 |
| 281 | For me, I would do this with a blank Excalibur | 对我来说,我会用一个空白的Excalibur做这个 |
| 282 | 或 TLDraw 棋盘。 | |
| 283 | I would write the final outcome at the top. | 我会写最后的结果 在顶端。 |
| 284 | Then I would draw the jobs, planner, | 然后我会画工作,计划员, |
| 285 | customer researcher, competitor researcher, distribution researcher, skeptic, merge, human | 顾客研究员 竞争者研究员 分销研究员 怀疑者 合并者 人 |
| 286 | approval. | 核准。 |
| 287 | Then I would draw the arrows. | 然后,我会画箭。 |
| 288 | The planner feeds the three researchers. | 计者供养三种研究者. |
| 289 | The researchers feed the skeptic. | 研究者供养疑者. |
| 290 | The skeptic feeds the merge. | 疑者供养合并. |
| 291 | The merge feeds the human decision, | 合并是人类的决定 |
| 292 | and that's enough. | 够了 |
| 293 | Now, once that works three times manually, then I would think about all | 现在,一旦它工作三次手动, 然后我会想到所有 |
| 294 | the tools. | 工具。 。 。 。 |
| 295 | The beginner version is a manual run with separate lanes, but the intermediate version | 初学者版本是手动运行,有单独的车道,但中间版本 |
| 296 | is cloud code, codecs, or a repo where each step writes files. | 是云码、解码器,或者每个步骤写文件的回波。 |
| 297 | The planner writes plan.md, | 计划员写计划 |
| 298 | the researcher writes customer.md, competitors.md, and distribution.md, and the skeptic writes | 研究者写客户Md,竞争者Md,发行者Md,怀疑者写作 |
| 299 | 评论.md. | |
| 300 | The merge step writes recommendation.md. | 合并步骤写作建议. md. |
| 301 | What's cool about that is it leaves a paper trail, | 酷的是它留下了纸迹 |
| 302 | and that's really nice. | 这是非常好的。 |
| 303 | You can see what happened. | 你可以看到发生了什么。 |
| 304 | You can compare versions, and you can | 你可以比较版本,你可以 |
| 305 | reuse the structure next week or a few weeks later. | 下周或几周后再使用这个结构. |
| 306 | Now, the advanced version is when you do | 现在,高级版本是 当你做 |
| 307 | 使用诸如LandGraph、Autogen Graphflow、N8n、Make.com或你自己的小脚本 | |
| 308 | actually orchestrate the graph. | 实际整理图。 |
| 309 | LandGraph is actually really useful when you want state | LandGraph 在您想要状态时确实有用 |
| 310 | checkpoints, persistence, human-in-the-loop approvals, and more reliable control over how | 检查站、坚持不懈、人与人之间的核准,以及更可靠地控制如何 |
| 311 | an agent workflow runs. | 代理工作流程运行。 |
| 312 | Then you have something like Autogen Graphflow, and that's useful when | 然后,你有类似自动图形流的东西, 这是有用的,当 |
| 313 | you want directed workflow with sequential steps, parallel steps, conditional branches and loops. | 您想要有顺序步骤、平行步骤、有条件分支和循环的定向工作流程。 |
| 314 | Tools like N8n, Make.com are useful when the graph touches everyday business systems like Slack, | N8n,Make.com等工具在图表触及Slack等日常商业系统时有用, |
| 315 | email, Airtable, or your CRM, but again, the tool is not the point. | 电子邮件, Airtable, 或您的CRM, 但同样, 工具不是重点 。 |
| 316 | The tool should come | 工具应该来 |
| 317 | after the workflow. | 工作流程结束后。 |
| 318 | If you automate a workflow, you do not understand. | 如果工作流程自动化,你不明白。 |
| 319 | You get a mess. | 你一团糟 |
| 320 | If you | 如果你们 |
| 321 | understand the workflow first, automation then becomes super obvious, and I can do a graph | 首先了解工作流程,然后自动化变得非常明显,我可以做一个图表 |
| 322 | engineering advanced tutorial if people are interested using things like LandGraph or | 如果人们有兴趣使用 LandGraph 或 |
| 323 | Claude Code, but for the purpose of this episode, I think we just want to get to level one and level | Claude Code, 但为了本集的目的, 我想我们只是想达到一级和一级 |
| 324 | two. | 两个 |
| 325 | Okay, so you now hopefully understand what graph engineering is at a high level, | 好吧,所以你现在希望 了解什么是图 工程在高层, |
| 326 | but how can you actually integrate this into your startup, into your business, to start | 但是,你怎样才能将它 真正融入你的启动, 与你的生意,开始 |
| 327 | making more money or creating better products or just generating a lot of value? | 赚更多钱还是创造更好的产品 或者只是创造很多价值? |
| 328 | The one that comes to mind first is customer support. | 首先想到的是客户支持。 |
| 329 | So a simple support graph | 那么简单的支持图 |
| 330 | might start by classifying the issue. | 不妨首先对问题进行分类。 |
| 331 | Is it billing? | 记帐吗? |
| 332 | Is it product confusing? | 是产品混淆? |
| 333 | Maybe it's a | 也许它是一个 |
| 334 | bug or cancellation risk or maybe it's something else. | 错误或取消 风险或可能是别的东西。 |
| 335 | Then the graph checks account context. | 然后图表检查账户上下文。 |
| 336 | So is it a new customer? | 所以这是新客户吗? |
| 337 | Are they high value? | 它们的价值很高吗? |
| 338 | Have they written in before? | 他们以前写过吗? |
| 339 | Are they frustrated? | 他们很沮丧吗? |
| 340 | Then it searches the docs or internal policies. | 然后它搜索文件 或内部政策。 |
| 341 | You might have like a whole wiki for your company, | 你可能会喜欢一个完整的wiki 对于你的公司, |
| 342 | maybe a notion board, maybe it goes and explores that. | 也许一个概念板, 也许它去探索。 |
| 343 | Then it drafts a reply. | 然后起草答复。 |
| 344 | {\fn黑体\fs22\bord1\shad0\3aHBE\4aH00\fscx67\fscy66\2cHFFFFFF\3cH808080}然后是检查器 | |
| 345 | reviews the reply for accuracy, tone, and risk. | 审查答复的准确性、语气和风险。 |
| 346 | Then a human approves anything involving refunds, | 然后一个人批准任何涉及退款的东西, |
| 347 | account changes, angry customers, legal risk, or promises that a company just might regret later. | 账户变化,愤怒的客户,法律风险, 或承诺一个公司 可能只是后悔。 |
| 348 | And that's the graph. | 这就是图。 |
| 349 | And it's better than saying AI answer the support ticket because the support | 这比说AI回答支持票好 因为支持 |
| 350 | ticket is not the real workflow. | 票不是真正的工作流程。 |
| 351 | The real workflow is understanding and researching and | 真正的工作流程是理解和研究以及 |
| 352 | drafting and checking and approving. | 起草、核对和批准。 |
| 353 | It's probably starting to click now. | 现在可能开始点击。 |
| 354 | Content creation is just | 内容创建只是 |
| 355 | another example that comes top of mind. | 另一个头顶上的例子。 |
| 356 | A content graph might start with research, then it creates | 内容图可以从研究开始,然后创造 |
| 357 | a thesis, then it finds examples, then it writes a hook, then it drafts a script, then a checker | 一个论文,然后找到例子,然后写一个钩子, 然后起草一个脚本,然后检查器 |
| 358 | asks whether the examples are specific, whether the pacing works, whether the hook earns intention | 询问实例是否具体, 速度是否有效, 钩子是否获得意图 |
| 359 | based on formats that are working, and whether the writing sounds like something like the person | 基于工作格式,以及写作是否听起来像一个人 |
| 360 | actually would say. | 实际上会说。 |
| 361 | Then the graph can branch into title ideas, thumbnail concepts, captions, | 然后图可以分化成标题构想,缩略图概念,标题, |
| 362 | b -roll,类似的东西。 | |
| 363 | And that's also closer to how a content lead, a real content lead that | 也更接近于一个内容如何引导, 一个真正的内容如何引导 |
| 364 | you would hire to help you create content would actually do. | 你会雇你来帮助你 创造内容会做。 |
| 365 | Another great example is coding. | 另一个伟大的例子是编码。 |
| 366 | A coding graph might start with a plan, then one agent edits the code, another reviews the diff, | 一个编码图可能从计划开始, 然后一个代理编辑代码, 另一个审查diff, |
| 367 | another runs tests, another checks the UI in a browser, another looks for edge cases, and then | 另一次运行测试,另一次在浏览器中检查UI,另一次查找边缘大小写,然后 |
| 368 | you have a human being actually approving the final pull request. | 你有个人 批准最后的拉动请求。 |
| 369 | And that's basically where | 这基本上就是 |
| 370 | all these AI coding tools are going. | 所有这些 AI 编码工具正在运行。 |
| 371 | The model writing the code is only one part of the workflow, | 写代码的模型只是工作流程的一部分 |
| 372 | and there's leverage in all the planning and testing and reviewing and inspecting and deciding | 在所有的规划和测试 以及审查和检查 和决定 |
| 373 | what is actually safe to ship. | 什么才是真正安全的船运。 |
| 374 | And that's actually an important point. | 这其实是一个很重要的问题。 |
| 375 | A big reason why graph | 一个大原因,为什么图 |
| 376 | engineering matters is it makes quality less dependent on someone remembering a perfect | 工程学上的问题在于 它使质量不那么依赖 一个人记得一个完美的 |
| 377 | prompt to ask their LLM. | 快速询问他们的LLM。 |
| 378 | It makes reviews way more consistent. | 这使审查更加一致。 |
| 379 | It makes delegation in general | 一般来说,它使代表团 |
| 380 | way cleaner. | 更干净 |
| 381 | It makes approval way more explicit. | 它使批准方式更加明确。 |
| 382 | It gives you a place to add tools and memory and | 它给你一个地方添加工具和记忆 |
| 383 | checks and permissions over time. | 检查和权限随时间推移。 |
| 384 | And it turns AI work from just like chat into this operating | 它把人工智能从聊天变成这个操作 |
| 385 | system. | 系统。 |
| 386 | And that really does feel like you're living in the future once you get to that place. | 感觉就像一旦你到了那个地方 你就活在将来 |
| 387 | Now, there is one mistake that I want to warn against, which is more agents don't automatically | 现在,有一个错误,我想警告, 更多的特工不会自动 |
| 388 | mean better output. | 意味着更好的输出。 |
| 389 | Sometimes actually more agents mean more noise. | 有时候更多的特工意味着更多的噪音。 |
| 390 | Sometimes it means | 有时意味着 |
| 391 | five AI workers confidently repeating the same wrong idea. | 5个人工智能员工自信地重复同样的错误想法. |
| 392 | Sometimes it means the system spends | 有时意味着系统花费 |
| 393 | more time coordinating than thinking. | 更多的时间协调 比思考。 |
| 394 | So the goal is not to make the biggest graph possible. | 因此目标不是让最大的图表成为可能. |
| 395 | I've | 我见过 |
| 396 | seen people on X go viral with these big, big graphs, but that's not the goal. | 我们看到X上的人用这些大块头的图表进行病毒传播,但这不是目标. |
| 397 | The goal is | 目标是 |
| 398 | and that's a really important distinction because a good graph should remove fake waiting and it | 这是一个非常重要的区别 因为一个好的图应该去除 假等待和它 |
| 399 | should separate workers from checkers. | 应把工人和支票员分开。 |
| 400 | And really it should be human approval where mistakes are | 在错误发生的地方 应该是人类的认可 |
| 401 | expensive and it should stop when the answer is good enough. | 当答案足够好的时候,它应该停止。 |
| 402 | It shouldn't need to continue. | 它不应该继续。 |
| 403 | And it should leave behind the useful state, the meaning notes, the evidence, the drafts, | 它应该留下有用的状态, 意义说明,证据,草稿, |
| 404 | the sources and the decision so that you can use it later. | 来源和决定,以便日后使用。 |
| 405 | And by the way, the last point is | 顺便说一句 最后一点是 |
| 406 | underrated because the real compounding value of graph engineering isn't just that one task | 被低估了 因为地图工程的真正复合价值 不仅仅是这个任务 |
| 407 | gets better. | 越来越好。 |
| 408 | It's that your work starts producing memory. | 你的工作开始产生记忆 |
| 409 | What do I mean by that? | 这是什么意思? |
| 410 | I mean that every | 我的意思是,每一个 |
| 411 | customer research graph creates better customer notes. | 客户研究图创造了更好的客户票据. |
| 412 | Every content graph creates better examples | 每个内容图都创造了更好的实例 |
| 413 | and audience insights. | 和观众的见解。 |
| 414 | Every support graph creates better product feedback. | 每个支持图都会产生更好的产品反馈. |
| 415 | And that's where the | 这就是... |
| 416 | becomes the moat because the graph produces the work, but it also produces the memory | 成为护城河,因为图 产生作品,但它也产生内存 |
| 417 | that makes the next graph smarter. | 这让下一个图更聪明。 |
| 418 | So it becomes this like asset for you. | 所以它就成了你的宝物 |
| 419 | So if you want to get | 所以,如果你想获得 |
| 420 | into graph engineering and you're like, how do I start? | 我怎么开始? |
| 421 | Here's a way to think about it. | 以此观想. |
| 422 | I would pick one workflow I already run with AI every week. | 我会选择一个工作流程 我已经运行每周AI。 |
| 423 | Maybe it's researching ideas or | 也许是在研究想法 |
| 424 | preparing podcast episodes, reviewing landing pages, analyzing customer feedback. | 准备播客节目,审查登陆页面,分析客户反馈. |
| 425 | Then I would write the final output in one sentence. | 然后,我将最后的输出写成一句话。 |
| 426 | For example, I want a one-page | 比如说,我要一页 |
| 427 | recommendation on whether this startup idea is worth testing. | 建议该启动想法是否值得测试。 |
| 428 | And then I would list the jobs | 然后我会列出工作 |
| 429 | a great human would do. | 一个伟大的人类会做的。 |
| 430 | They would clarify the question. | 他们将澄清这个问题。 |
| 431 | They would research the customers. | 他们会调查客户 |
| 432 | They would research competitors. | 他们会研究竞争对手 |
| 433 | They would look for distribution. | 他们会寻找分配。 |
| 434 | They would look for risks. | 他们会寻找风险。 |
| 435 | They | 他们 |
| 436 | would draw arrows where the work actually depends on another step. | 将绘制箭头 工作实际上取决于另一步骤。 |
| 437 | So what do I mean by that? | 我这是什么意思? |
| 438 | Customer research and competitor research could happen at the same time. | 客户研究和竞争者研究可以同时进行。 |
| 439 | The skeptic needs the | 怀疑论者需要 |
| 440 | research before it can check it. | 在检查之前先研究一下 |
| 441 | And the final recommendation needs the skeptic pass before it | 最后的建议需要先通过怀疑者才能通过 |
| 442 | can merge the evidence. | 可以合并证据。 |
| 443 | Then I would add one human gate before the expensive decision. | 然后,我会在昂贵的决定之前增加一个人类大门。 |
| 444 | If the output | 如果输出 |
| 445 | is a private memo, maybe the human gate is light. | 这是私人备忘录,也许人类的大门是轻的。 |
| 446 | If the output is a customer email, a public post, | 如果输出是客户电子邮件,公共邮箱, |
| 447 | code deploys, a refund or anything touching production data, you got to have a human gate | 代码部署,退款或者任何触摸到生产数据的东西, 你必须有一个人类大门 |
| 448 | that's stricter. | 这更严格。 |
| 449 | Then I would run it manually once. | 然后我会手动操作一次。 |
| 450 | This is the whole first rep that we want to get | 这是第一个代表 我们想得到 |
| 451 | good at. | 不错 |
| 452 | You don't have to create this giant automation project. | 你不必创建这个巨大的自动化项目. |
| 453 | Just create the jobs and | 只要创造工作机会 |
| 454 | arrows. | 箭头 |
| 455 | After you do this once, you start seeing AI work differently. | 做一次之后,你开始看到AI的工作方式不同. |
| 456 | Because you're not thinking | 因为你没有在想 |
| 457 | about like, okay, I need to do the most perfect prompt ever. | 有关,好吧,我需要做 最完美的快速,有史以来。 |
| 458 | What is the perfect prompt for this | 还有什么好及时的? |
| 459 | task I'm trying to do? | 任务我试图做? |
| 460 | You start thinking about, okay, what's the most perfect workflow for this? | 你开始思考, 好吧, 什么是最完美的工作流程吗? |
| 461 | And then you start designing a path that produces that answer. | 然后,你开始设计一个路径 产生答案。 |
| 462 | That's why I think | 所以我觉得 |
| 463 | graph engineering in general is a concept that is worth paying attention to. | 图表工程一般是一个值得注意的概念. |
| 464 | It's really like | 这真的像 |
| 465 | the next logical step after prompting. | 提示后的下一个逻辑步骤。 |
| 466 | And I think the people who get the most out of AI | 我觉得那些从AI里得到最多的人 |
| 467 | will be the people who know how to break down, work into the right pieces, give each piece the | 将会是那些知道如何分解的人, 工作在正确的片段, 给每一片 |
| 468 | right context, check the output, and keep the human in the right place. | 右上下文,检查输出, 并保持人的位置正确。 |
| 469 | So now that we're | 所以,现在我们 |
| 470 | towards the end of the episode, here's what I would do to try to learn this. | 在剧情即将结束时, 我将努力学习这个。 |
| 471 | I would pick | 我会选 |
| 472 | one workflow you already run, draw those jobs and arrows, delete the fake waiting, run the | 一个您已经运行的工作流程, 绘制这些任务和箭头, 删除假等待, 运行 |
| 473 | independent jobs in parallel, add a skeptic, merge the survivors, approve the final step yourself, | 独立工作并行 增加怀疑 合并幸存者 自己批准最后一步 |
| 474 | and there you have it. | 你拿着 |
| 475 | That'll be your first graph. | 这是你的第一个图表。 |
| 476 | And once you have one graph that works, | 一旦你有了一张有效的图表 |
| 477 | you're not just prompting AI anymore. | 你不只是提醒AI了。 |
| 478 | You're managing AI work. | 你在管理人工智能 |
| 479 | It's sort of this like next | 碞钩妓 |
| 480 | level in being an agent manager and really just like stepping yourself into this new world, | 成为代理经理,就像踏入新世界一样 |
| 481 | deep into this new world where you're getting the most out of AI to build out your dreams, | 深入到这个新世界里 你从AI那里得到最多 来创造你的梦想 |
| 482 | to take ideas and put them out there and getting, you know, something I just think a lot about now | 带着想法,把它们放在那里, 得到的东西,你知道, 我只是想了很多现在 |
| 483 | is just like, how do I get the most out of these platforms? | 我怎么才能从这些平台里得到最多? |
| 484 | And graph engineering is just a concept | 而图表工程只是一个概念 |
| 485 | that helps you think about that. | 帮助你思考。 |
| 486 | So there you have it, folks. | 所以,你们有它,伙计们。 |
| 487 | Graph engineering clearly explained. | 图表工程有清楚的解释。 |
| 488 | Hope that it got your creative juices flowing. | 希望它让你的创造力流畅 |
| 489 | Hope it's been helpful. | 希望有帮助 |
| 490 | My name is Greg Eisenberg. | 我叫格雷格・艾森伯格 |
| 491 | I'm the host of the Startup Ideas Podcast. | 我是"创业创意"的主持人 |
| 492 | For more, | 为了更多 |
| 493 | you know, like, comment, and subscribe. | 你知道,像,评论和订阅。 |
| 494 | 跟着我们Spotify和苹果。 | |
| 495 | And, you know, I feel | 而且,你知道,我觉得 |
| 496 | grateful that you're here, that I'm able to teach you, give you these concepts. | 感谢你们在这里, 我能够教你, 给你这些概念。 |
| 497 | And I just can't | 而我却无法 |
| 498 | wait to see what you build. | 等着看你们造什么 |
| 499 | I'm rooting for you. | 我为汝为根. |
| 500 | Have a creative day and I'll see you next time. | 祝你有创意,下次见 |
该视频共有字幕 500 条。解锁更多字幕为会员功能，请移动到 价格