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「Joining Ought」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

Starting at the end of August, I’m joining Ought as their first product designer. Ought is a non-profit research lab. They’re exploring how natural language processing (NLP) and machine learning (ML) tools can improve researcher workflows. Specifically, how these new tools can help people with open-ended reasoning – thinking through problems that don’t have simple, clear-cut answers.

Their main product at the moment is Elicit – an AI assistant for academic and professional researchers. It uses NLP to find research papers, synthesise them, and extract research questions, evidence, and arguments from them. They’re currently focused on helping people do literature reviews. But the plan is to expand Elicit’s capacities to help with the whole research process.

默认界面在强化什么

I got access to the alpha version of Elicit in July of 2021 5ya and was immediately hooked. Even though I’m an amateur researcher Meaning I do it as part of my job as a designer and writer, but in a rather a naive way compared to anyone writing a PhD. , I still spend a good chunk of time hunting down and reading academic publications.

I found the results were on par with what Google Scholar or Semantic Scholar would turn up. But the Elicit results show why it returns certain papers – each paper has a GPT-3 generated summary that tries to answer the original question. It’s a small difference, but a huge help when you’re drowning in PDFs and trying to quickly find the right ones to read. As with all things generated by automated systems, these results aren’t meant to be swallowed whole wi

可以怎么改

I’ll stop with the sales pitch now. 😉 Elicit is free and you can try it out if you want to see for yourself.

I’m also excited to explore what’s possible beyond summarising papers. In Programmable Notes I discussed what it might look like to add AI agents into personal notes and knowledge management systems. Many of Elicit’s secondary workflows point in that direction. You can generate research questions, rephrase ideas, or explore chains of reasoning .

落到产品里的动作

Suffice to say, there’s a lot to dive into. Language models and neural networks are all relatively new. GPT-3 is barely 2 years old. We don’t have many established design patterns or canonical interfaces for this stuff yet.

I am also new to this space. I haven’t worked on any ML and NLP projects yet so I have a fat reading list to work through. Like most people, I’ve heard plenty of cultural narratives around the nebulous concept of “AI”. Anthropologists like Nick Seaver , danah boyd , and Genevieve Bell have given me a good critical lens on the space. But I’ve never looked into the details of how these systems work.

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