Maggie Appleton 写设计、人类学和智能体界面时，习惯先画清楚隐喻，再谈工具。把「Common Misconceptions in AI」整理成可阅读的中文笔记：问题在哪、界面默认会把人带去哪、落地时该改什么。原站导航、评论回响和广告已去掉。

问题怎么被看见

A roboticist breaks down common misconceptions about what’s hard and easy in robotics. A response to everyone asking “can’t we just stick a large language model into its brain to make it more capable?”

Contrary to the assumptions of many people, making robots perceive and move in the world in the way humans can turns out to be an extraordinarily hard problem to solve. While seemingly “hard” problems like scoring well on intelligence tests, winning at chess, and acing the GMAT turn out to be much easier.

默认界面在强化什么

Everyone thought it would be extremely hard and computationally expensive to teach computers language, and easy to teach them to identify objects visually. The opposite turned out to be true. This is known as Moravec’s Paradox .

Especially liked the ending where Dan explores why people are so resistant to the idea picking up a cup is more complex than solving logic puzzles. Partly anthropocentrism; humans are special because we can do higher order thinking. Any lowly animal can sense the world and move through it. Partly social class bias; people who work manual labour jobs using their bodies are less valued then people who sit still using their intellect to solve problems.

可以怎么改

先写出默认交互会鼓励什么行为，再决定要不要用聊天框当万能入口。没有对照的界面改动，很容易只是换皮。

落到产品里的动作

先写出默认交互会鼓励什么行为，再决定要不要用聊天框当万能入口。没有对照的界面改动，很容易只是换皮。

落地时建议先做的 5 件事

1. 先写清这个工具在强化思考还是在替人思考。
2. 默认交互不要只会谄媚：该追问、该给反例、该要求证据。
3. 智能体界面要露出推理步骤，而不是只给一个光滑答案。
4. 知识发布用可生长的笔记，而不是一次性营销长文。
5. 改产品前先画隐喻：用户以为自己在做什么，系统实际在做什么。

和智能体产品怎么接

龙虾PRO做 OpenClaw 落地时，最该从这类笔记里拿走的是「别把聊天框当唯一界面」。数字员工要能追问、能验证、能把过程摊开，而不是只负责说好话。

本文侧重智能体生命周期与技能边界。若你要动手验证，优先用 Playground 做小任务压测；权限与托管再单独规划，避免「先装一堆再治理」。 相关：Playground · 创建智能体 · 学习中心