引言

我的理解

本课以“Gmail 邮件优先级分类”为贯穿全模块的端到端案例,正式介绍用 GenAI 构建工具时的完整生命周期。核心方法论包含四个执行步骤(识别瓶颈、技术可行性、智能可行性、构建)以及四项配套学习(有选择地委托、文档管理、评估机制、风险管理与可观测性),这些能力共同构成一套为 GenAI 不确定性量身定制的防御性方法论。令人意外的是,这套方法与优秀人类管理者的思维框架高度吻合,暗示 AI 管理本质上是人员管理的延伸,而不只是一种技术技能。

相关链接


原文

Lesson 27 of 68 引言 / Introduction

通过上周的学习,我们已经具备了 Builder 思维。你可能已经将它应用到日常工作中,并从中受益。在本模块中,我们将通过另一个端到端的案例研究,更深入地探讨其中的常见细节与推荐实践。

我们将以“将邮件优先级排序任务委托给 GenAI”为例。这样我们就不必频繁查看邮件,AI 会自动提醒我们哪些邮件重要或紧急,让我们能够专注于真正需要判断力和创造力的任务,同时不必担心错过任何信息。我们会以 Gmail 作为一个抽象示例,演示用 GenAI 构建工具的完整生命周期。这有助于展示许多实用技巧与方法,它们同样适用于 Slack、Zendesk 等其他工具。

构建过程应当在 AI 之外完成。在第一节课中,我们将学习如何分析自身的工作流程,识别出适合委托给 GenAI 的任务,同时也帮助识别潜在的风险点。接着,我们会讨论应对这些风险在技术上的可行性。然后,我们将介绍如何配置 GPT API 以便以编程方式调用 GenAI。我们需要从“智能能力”的角度验证 GenAI 是否能成功完成我们的任务。为了管控风险,我们会建立一种机制,让 GenAI 能够尽早、快速地失败。如果遵循 GenAI 的最佳实践,真正的构建阶段会非常顺畅。最后,我们会建立可观测性,确保任何失败都能被及时发现,避免错误悄无声息地发生。

我们之所以采用一套专为 GenAI 量身定制的方法论,是因为用 GenAI 构建本身就带有独特的挑战。这要求我们在整个生命周期中具备特定的思维方式和技能组合,以应对其中的细微差别。在本模块中,我们将学习以下内容:

有选择地委托给 AI:先从优化工作流程入手,而不是不加区分地使用 AI。AI 并不总是提升效率的最佳方案。

风险管理:与传统工具不同,GenAI 可能达到预期,也可能不达预期。承认失败并调整方向是可以接受的,往往也是必要的。有一些实践方法可以帮助我们最大限度地减少时间浪费。

评估机制:在执行之前先设定成功标准。在不同任务中评估 AI 的表现,从而建立属于你自己的经验和判断。

文档管理:当我们已有的内容能用精确的文档(如代码或代码库)来描述时,GenAI 的表现会非常出色。为你构建的工具持续维护这些文档,会随时间产生复利式的效果。

可观测性:工具投入使用后,工作并未结束,对自动化工具尤其如此。至少,你需要知道什么时候出现了故障。错误的结果比没有结果糟糕得多。

如果我们审视上述要点,会发现它们与一位优秀的人类团队管理者的思考方式高度契合。在本模块的最后,我们会探讨这种有趣的相似之处,并以一个关键的可执行洞察作为收尾:提升 GenAI 能力最有效的方式,或许并不是只钻研数学和机器学习,而是去学习如何做好人员管理。

English Original

With last week’s learning, we now have the Builder mindset. You’ve probably even applied it to your daily job and benefited from it. In this module, we will examine another end-to-end case study to delve deeper into common nuances and recommended practices.

We will use an example of delegating the task of email prioritization to GenAI. This way, we can avoid constantly checking emails. Instead, the AI will automatically notify us of important or urgent emails, allowing us to focus on tasks that require our judgment and creativity without worrying about missing anything. We will use Gmail as an abstract example to demonstrate the entire life cycle of building something with GenAI. This will help illustrate many practical tricks and skills that can also be applied to tools like Slack, Zendesk, and others.

The building process should happen outside of AI. In the first lesson, we will learn how to analyze our workflow and identify tasks suitable for delegation to GenAI. This will also help identify potential risk points. We will then discuss the technical feasibility of addressing these risks. Next, we will cover the process of setting up GPT APIs for programmatic invocation of GenAI. We need to verify whether GenAI can successfully accomplish our tasks from an intelligence perspective. To manage risks, we will establish a mechanism for GenAI to fail quickly and early. The actual building phase will be straightforward if we follow best practices with GenAI. Finally, we will set up observability to ensure any failures are detected, preventing unnoticed errors.

We used a unique methodology tailored for GenAI because building with GenAI presents its own unique challenges. This requires specific mindsets and skill sets throughout the entire life cycle to address these nuances. In this module, we will learn about:

Selective Delegation to AI: Start by optimizing workflow rather than using AI indiscriminately. AI is not always the best solution to boost productivity.

Risk Management: Unlike traditional tools, GenAI may or may not work as expected. Admitting failure and iterating on the direction is acceptable and often necessary. There are practices to minimize wasted time.

Assessment Mechanism: Set success criteria before execution. Evaluate AI performance in various tasks to build your own experience and opinions.

Document Management: GenAI performs well when precise documents, such as code or libraries, describe what we already have. Maintaining these documents for the tools you build will have a compounded effect over time.

Observability: The process doesn’t end when the tool is in use, especially for automation tools. At a minimum, you need to know when something fails. A wrong result is much worse than no result.

If we examine the points above, it’s clear they closely mirror how a human team manager thinks. At the end of this module, we will explore this interesting resemblance and conclude with a key actionable insight: the most effective way to enhance your GenAI skills may actually be through learning people management rather than focusing solely on mathematics and machine learning.