Danny Quah

Danny Quah

Li Ka Shing Professor in Economics

AI inversion. I asked my local LLM to introspect and join up what I’d been thinking the last four years

2 minute read

Published:

“AI inversion. I asked my local LLM to introspect and join up what I’d been thinking the last four years”

by
Danny Quah
Feb 2026

This past weekend I looked at my LLMs and decided that for a while I wasn’t going to be asking what the world says or knows. Instead, I wanted to see the connections in and clusters of ideas, common turns of phrase, and formal parallels in statistical models and mathematical structures in and across the models of social mobility and of world order that I have been developing. The source material had to be clearly circumscribed and controlled: notes I keep of half-finished conjectures and ideas from what I’ve read; YouTube and podcast recordings of panels, roundtables, and fireside chats I’ve done; markdown and latex source of papers, blogs, and commentaries I’ve written.

I wanted the ideational connections across these mathematical and conceptual models updated in real time, and I didn’t want this source material uploaded to the cloud. I wanted to query locally and dynamically my second brain, not just look at its connections graphically, which Obsidian had already long allowed me to do.

So I rigged up on my MacBook Air an ollama runner, and downloaded onto it a Chat Model and a RAG Embedding Model, compact enough to run in 16Gb RAM. Then, because everything happens locally, there are no API or subscription charges, and I know any conclusion I draw comes only from thoughts I had myself.

It works. I got connections, gaps, blindspots, suggestions.

In the 1m02s animation, each of the nearly 6,000 nodes is a document, URL, or other link; and the size and links represent ideational footprint and connection strength.

(If I’d known ahead of time the below was all I had to do, it would have taken only 5mins to get up and running:

curl -fsSL https://ollama.com/install.sh | sh  
ollama pull llama3:8b-insgtruct-q4_K_M  
ollama pull nomic-embed-text

Obsidian Copilot plugin:

  • Self-hosted ollama API http://localhost:11424; ignore all other API key requests
  • Chat model llama3:8b-insgtruct-q4_K_M; embedding model nomic-embed-text
  • For API key of both chat and embedding models, don’t leave blank but put in any string (like x or ollama).
  • In tab QA, set embedding model to nomic-embed-text (Ollama)
  • Use vault QA (free)

You May Also Enjoy