研究工作常散落在論文、實驗紀錄和聯絡資料裡,AI 助理如果能在需要時找到合適資訊,就能提醒你以前做過的實驗,或建議你重新聯絡某個人。
- 作者把這套做法稱為「知識工廠」,用多個 LLM Wiki 整理研究論文和實驗結果,並以 Markdown 檔案和圖譜呈現。他建議先和 AI 助理一起使用、慢慢調整知識庫,不必一開始就訂下複雜規則。
- 整理好的知識也能支援評測、內容製作等工作。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
Recommended reading on building a "knowledge factory".
The biggest transformation in my agent setup recently is feeding it knowledge at the right time.
This compounds in insane ways with proactive agents that can connect bits of knowledge over time.
It's such a magical experience when your agent reminds you of an experiment or suggests you reach back out to a connection.
There are no standards for how you build a knowledge factory. For instance, I use graphs of MD files. MD files are versatile and are the preferred language of LLM agents, and graphs provide structure, which turns out to help agents a ton. OKF is an interesting format for capturing semantics better, but I agree you don't want to be overprescriptive here.
I have built several LLM Wikis (research papers, experimental results, etc.) that together make up my knowledge factory.
In terms of building it, what I have found works best is interacting with your agents and co-evolving the knowledge factory as you go. No need to overcomplicate things initially.
As a researcher, I've built other forms, like an eval factory and a content factory, that build on top of this knowledge factory; they are all key to the software factory. Okay, I know I used the word "factory" too much, but you get the point. This compounds.