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 intermingle’

Or: starting from “Austin’s homepage”

Start with landing pages for each person

Big data collection & viz project: Show Austin’s texts, emails, comments, posts

Then: show Austin’s view of individual folks eg Constance

Which comments & posts he’s seeing, the mental model of Constance in Austin’s head, top level “what work she does” and finer “these quirks are memorable”
Key point: Austin’s view of Constance differs from Rachel’s view of Constance

Then: show Austin’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 “head”/follow graph

Project is about finding the vector in idea-space that is Austin

Inspirations

Personal, LLM-augmented sites

Gwern’s website

Gwern Branwen Nenex: A Neural Personal Wiki Idea
Gwern Branwen Towards Benchmarking LLM Diversity & Creativity

‣

Virtual Comments

Simon Willison’s Weblog

Really like the mix of short quotes, long summaries

Character.ai
Tyler Cowen’s book

Freely mixing

Act I:
Act I: Exploring emergent behavior from multi-AI, multi-human interaction
Subreddit Simulator
…? Should be more of this

Finetuning characters

Sarah Constantin Fine-Tuning LLMs For Style Transfer
Alexander Wales:

Advanced prompting guide on Replicate: https://replicate.com/docs/guides/language-models/advanced-prompting

Concepts

Worth the Candle: Nachless (avatar formed from writings)
“The Jesus that you simulate in your head is real”, from Emmett Shear

Other notes

Ethan Mollick 15 Times to use AI, and 5 Not to

Dev notes

‣

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 don’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 she’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?

You are simulating how Scott Alexander would write an Astral Codex Ten post about ADHD. Use his characteristic style: - Start with an interesting observation or paradox - Build through multiple examples and studies - Use clever analogies and occasional humor - Include careful probabilistic reasoning - End with a surprising synthesis Here are examples of his style: [Example 1 - from his post "Ketamine Research In A New Light"] "A few weeks ago, Nature published a bombshell study showing that ketamine’s antidepressant effects were actually caused by a metabolite, 2S,6S;2R,6R-hydroxynorketamine (don’t worry about the name; within ten years it’ll be called JOYVIVA™®© and you’ll catch yourself humming advertising jingles about it in the shower). Unlike ketamine, which is addictive and produces scary dissociative experiences, the metabolite is pretty safe. This is a big deal clinically, because it makes it easier and safer to prescribe to depressed people. It’s also a big deal scientifically. Ketamine is a strong NMDA receptor antagonist; the metabolite is an AMPA agonist – they have different mechanisms of action. Knowing the real story behind why ketamine works will hopefully speed efforts to understand the nature of depression. But I’m also interested in it from another angle. For the last ten years, everyone has been excited about ketamine. In a field that gets mocked for not having any really useful clinical discoveries in the last thirty years, ketamine was proof that progress was possible. It was the Exciting New Thing that everybody wanted to do research about. Given the whole replication crisis thing, I wondered. You’ve got a community of people who think that NMDA antagonism and dissociation are somehow related to depression. If the latest study is true, all that was false. This is good; science is supposed to be self-correcting. But what about before it self-corrected? Did researchers virtuously say “I know the paradigm says NMDA is essential to depression, and nobody’s come up with a better idea yet, but there are some troubling inconsistencies in that picture”? Or did they tinker with their studies until they got the results they expected, then triumphantly declare that they had confirmed the dominant paradigm was right about everything all along? This is too complicated an issue for me to be really sure, but overall the picture I found was mixed." [Example 1 - from his post "Depression"] "The short version: Depression has many possible causes, including stressful life events and biological problems like inflammation and hormone imbalances. Any given case of depression might be due to some of these causes and not others. Some people have long-standing mild depression (dysthymia), but more often depression comes in episodes; these usually go away on their own in 3 – 12 months. If you want your depression to go away sooner, there are lots of things you can do: life changes, diet, exercise, therapy, supplements, medications, and high-tech options like transcranial magnetic stimulation. Different options work for different people. If you try one option and it doesn’t work, move on to another. Your doctor will work with you to start with safe and easy options, then escalate to stronger and more difficult ones if those don’t work; the whole process will involve a lot of trial and error but should result in something that works effectively and consistently for you. Some options might seem too hard at the beginning (eg if you’re very depressed you might have trouble maintaining an exercise routine), but might get easier after you’ve tried other options (eg you start with medication, that makes you feel a little better, and then you can try exercising). Once you find something that works for you, continue it for at least six months to get through your current depressive episode, then consider whether you want to try going without it, or whether you want to continue it indefinitely."" Now, write the opening section of a Scott Alexander post titled "ADHD".