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「Squish Meets Structure」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。

Video Recording

This talk is about designing products with language models – how to think about models, understanding how vectors and embeddings work, and some priciples for designing interfaces and systems that don’t drown your users in cognitive overload.

First given online for Smashing Conference AI back in mid- 2023 3ya , and then in person in Freiberg and Antwerp. Here’s some snaps from the Freiberg edition:

Slides and Transcript

The front entrance of the conference – Source: Fabian Kaegy Crowd shot with one of my slides – Source: Marc Thiele Video Recording Talk runs from 0:00 – 34:30, Q&A from 34:30 – end

This talk is going to be about designing products with language models.

值得单独记下的观察

• Useful ways to think about language models
• The tension between traditional programming and LMs
• How and why I think we should design models to be tiny, specialised reasoning engines
• Fine-tuning is just training it on a small data set and telling to model to pay more attention to this data than the rest of its training data.
• Reinforcement learning is when you get humans to score the output and feed that back into the model so it learns what kind of answers the humans want. OpenAI heavily used this technique to improve ChatGPT.
• The amount of reinforcement learning done on the base model will affect how constrained the outputs are
• Little or no prompt engineering will give you more open-ended responses, while highly constrained prompts with specific examples will give you structured answers
• extracting structured data from long text

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。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源