For a while now, I've wanted something like Connected Papers for LessWrong & the broader rationalist community.
I built a working prototype that has, so far, been pretty useful to me. You can find it at: situate.info/rational
Usage
You give it a LessWrong or Substack URL and receive a list of the top 100 similar documents from the database.
Notes:
* It was updated on 4/19, newer posts won't work.
* I do not currently have the list of scraped Substacks published (should add it to the website) but it includes Astral Codex Ten, Overcoming Bias, and a bunch of other rationalist-adjacent publications.
How it works
I scraped ~372k documents:
* LessWrong: 36,220 posts, 6,329 top level shortforms, 289,721 comments. Filtered by karma ≥ 5.
* Substack: 40,289 posts across 173 publications.
Each document was split into ~360-token chunks (heading-aware, sliding window, or whole-doc for short pieces), embedded with Voyage-3.5 (1024-d), then averaged into a single doc vector. Similarity is calculated via cosine distance for a total of 100 matches per query.
What's next
I'm excited to receive feedback/ideas from you all, for further iteration.
I have a personal setup of this tool that allows for open-ended custom queries (embedded on-demand), and BGE reranking for slightly improved accuracy. At the moment, this would be too expensive/complicated for me to implement at scale, but I do currently plan to build a polished setup for the ingestion -> chunking & embedding -> search pipeline, that I can open source for others to use.
A semantic search on LessWrong alone would probably be an easy project for the Lightcone team, so I expect that specific function to eventually become obsolete. However, I plan to keep ingesting information from many different corners of the internet, so I think building an effective third party system is still useful.
In general, I'm interested in working on interfaces that improve A) individual navigation and synthesis of information and B)