How I Work · 2026-08-16 · 12 分钟

AI slop 与 slop fixing

核心摘要

人们与AI合作时往往陷入“slop fixing”(反复修改AI低质量输出),导致效率不升反降。核心判断:改用“architecting”方法——在要求AI产出前充分沟通背景、目标和约束——能让第一版就达到80%-90%的准确度,大幅减少后续修复时间。

干货提炼

主题一:AI slop 与 slop fixing

  • AI slop 是表面好看但无实质内容的输出。 证据:主持人描述“it kind of looks good on the surface, but when you get into it, there is absolutely no substance”,且包含大量泛泛之谈(generalities)。
  • slop fixing 是用户反复与AI争论、修改低质量输出的过程,消耗大量时间。 证据:研究显示人们每周因此浪费近一天时间(“we spend nearly one day a week slop fixing”)。典型表现是“go back and forth and back and forth with the AI, arguing with it”,或干脆放弃接手自己修改。

主题二:Gunslinging 与 Architecting 的对比

  • Gunslinging(枪手式)是投入极少前置信息、直接要求AI产出,然后大量修改。 证据:类比“shooting from the hip”,只给简单提示如“build me a house”,之后反复抱怨房子不符合要求。前置投入小,后期修复成本高,且“fix one thing breaks something else”。
  • Architecting(架构师式)是花更多时间在前置讨论上,让AI理解目标、背景、约束后再产出。 证据:类比找建筑师,先讨论家庭需求(四口之家、狗、丈夫需要工作间)再设计,而非直接说“造个房子”。前置投入稍多,但第一版产出更精准,后续只需少量润色。

主题三:Architecting 的实践方法

  • 具体做法:像与员工或顾问合作一样,先展开对话,共同规划蓝图。 证据:以“为新员工(营销主管)制定入职计划”为例,不应直接说“写个入职计划”,而应说明岗位是数字营销还是传统营销、公司目标是拓展国际市场等,然后与AI讨论最佳实践,再让AI产出最终版本。
  • 关键原则:将努力前移(effort bias),让AI在第一版就接近你的需求。 证据:主持人总结“with AR it's just a little bit more time upfront making sure that it knows what you're really trying to achieve”,这样“version one is more likely to nail what you need... probably going to be eighty or ninety percent of the way there”。

高光金句

  • ❝ It's funny because it kind of looks good on the surface, but when you get into it, there is absolutely no substance. ❞
  • ❝ With Gunsling, you've got very amount of a small amount of effort at the front, but then you've got a lot of effort down the track because it's then fixing, and often when you fix one things, it breaks to something else and then it becomes quite problematic. ❞

提及资源

  • 本集暂无提及外部资源

在 Readio 中打开本集

全文检索、逐字稿阅读、就这一集的内容直接向 AI 追问——读完摘要还想深挖的话。

Podup iOS 应用下载二维码扫码下载 iOS 应用