šŸ’¾

simulation capture

What

Simulate a bunch of folks (myself, friends, acquaintances, writers like Scott Alexander), and be able to talk with them. Then let them talk to each other.

Aka: Simulation capture; LLM echo chamber

Why

  • Unlock the inherent bottleneck in individual humans only having 24h a day (like Amaryllis clones from Worth the Candle)
  • Reduce communication costs. E.g. Before drafting a message to Scott, can talk with his clone first
  • Formalize the nature of internet text communications/Manifest invite pathway where you get to know your idols via what they write
  • Help myself remember folks, stay in touch, outsource relationships to a network

How

  • Collect data
    • Ideally: Point a website/blog/corpus/Twitter/LW into the thing and it pulls out a bunch of text examples
    • Or: Starting point of personal notes? (like, my Notion?)
  • Turn into LLM: finetuned model (or a giant prompt to start)
  • Let people chat with them, Character.AI style
  • Let them intermingleā€™

Or: starting from ā€œAustinā€™s homepageā€

  • Start with landing pages for each person
    • Big data collection & viz project: Show Austinā€™s texts, emails, comments, posts
  • Then: show Austinā€™s view of individual folks eg Constance
    • Which comments & posts heā€™s seeing, the mental model of Constance in Austinā€™s head, top level ā€œwhat work she doesā€ and finer ā€œthese quirks are memorableā€
    • Key point: Austinā€™s view of Constance differs from Rachelā€™s view of Constance
  • Then: show Austinā€™s info feed
    • Permit other humans to read from, but not edit
    • Or: show the network that is just people you know interacting with each other
      • e.g. seeing Scott Sumner & Hanania talk is much more relevant than upvoted randos. (Twitter algo does this, I think).

Consider

  • Essential nature of a person is more than just what they write?
    • Relationships to others
    • Interests, tastes, what they read
    • Physical characteristics, in-person charisma
      • (famously, words are only 30% of in person communication or wtv)
  • Where will an LLM fall short? How to augment/fix that?
    • Initiative
    • Long term planning
    • Or: How to lean into strengths of LLM (eg always available)
  • Difference contexts, writing styles between private notetaking, public comments, DMs
  • Consider memory as a bottleneck? Any agent has limited bandwidth
    • E.g. maybe agents can only keep 150 folks in their ā€œheadā€/follow graph
  • Project is about finding the vector in idea-space that is Austin

Inspirations

Personal, LLM-augmented sites

Freely mixing

Finetuning characters

Concepts

  • Worth the Candle: Nachless (avatar formed from writings)
  • ā€œThe Jesus that you simulate in your head is realā€, from Emmett Shear

Other notes

Dev notes

ā€£
Experimented with a giant prompt that Claude helped write, to a few different models via OpenRouter
  • Claude Sonnet 3.5 is also pretty good at simulating Scott, at least?
  • Got sidetracked into reading an ADS post with Alex Berger
    • (one hazard of working on this kind of project)
    • But also: goal is to be able to extract the gestalt of Alex Berger (and ADS), so that I donā€™t have to go read articles on my own?
      • Also: an important thing I do is, when I read something, connect it to something else in my head. Like just now I sent another ADS article to Elizabeth van Nostrand, when reading one thing sparked a connection to something else
        • The act of sending that message is a bit like, Austin simulates ADS while reading, then simulates Elizabeth (that sheā€™d like the link), then forwards the connection
  • Another thing that writers do is cite/hyperlink to their own work a lot
  • Also: would like some kind of research notebook, like Notion with LLMs but good?
    • Or Jupyter?