Micah Stubbs' Weblog

Simulating 6,245 real people on one machine

20th August 2026

Last night I went to a simulation build night built around two papers that simulate absurdly large artificial societies: Light Society (a billion LLM agents) and MatrAIx (8.3 billion persona records). Both are impressive. Neither is something you can run at home.

So with Yvonne Chen and Jake Schwartz, I built popsim — an attempt to reproduce the load-bearing elements of both papers on a single machine, over one public dataset, in an evening. Everything runs in under 300 MB of RAM.

Finding a dataset that still exists

The first surprise was how thoroughly the tweet-dataset landscape has collapsed. Since X shut down free API access in 2023, sharing tweet-ID lists for “rehydration” is effectively dead — the IDs are out there, but turning them back into text requires an enterprise contract.

What you want now is full-text corpora, and the best one for persona work turns out to be seventeen years old: Sentiment140, 1.6 million labeled tweets from 2009. Measured on distinct users with at least 20 tweets each — enough history to ground a persona — it beats the runner-up by an order of magnitude: 6,245 qualifying users against the Community Archive’s 363 accounts.

Five experiments, one honest negative result

The interactive explainer has the full story. The short version:

Persona mining. Every user with ≥20 tweets became a persona card — sentiment disposition, verbosity, mention rate, peak posting hour, vocabulary. MatrAIx’s human-grounded personas at hackathon scale.

A scale ladder. Light Society’s trick, one node: full LLM agents for the 927 heaviest users, cheap surrogates for 5,318, statistical samplers for the rest. 702.7× cheaper per tick than all-LLM.

Diffusion vs. ground truth. The cohort has a real social graph — 31,515 @-mention edges. I ran DeGroot opinion diffusion over it and asked whether that predicts second-half sentiment better than assuming people stay themselves. It doesn’t: r = 0.54 against r = 0.73 for plain persistence. Disposition beats diffusion, at least on a 48-day horizon.

Parasocial structure. Strong mutual bonds (0.6% of pairs) and strong one-way attachments (0.7%) are almost exactly balanced, and 12.1% of mention volume flows to accounts that never reply. That broadcast tier costs almost nothing to simulate — it’s one-directional by definition.

A benchmark. Per-user temporal 80/20 split, sentiment adherence on 42,998 held-out tweets. The zero-parameter baseline gets 63.1%; MatrAIx’s 91.5% frames the gap. Section 06 has a copy-paste prompt that points any agentic LLM at the repo to run its own research loop. I ran it myself: TF-IDF plus per-user priors reaches 82.0%, then plateaus. The remaining ten points probably need a real language model.

The part I didn’t expect

Both surprises push the same way. Diffusion failed informatively — who you are predicts your behavior better than who you talk to — and the parasocial analysis says most of the attention economy is one-directional anyway. If you’re simulating a population on a budget, spend it on the personas.

Code is in pop-world-model-sim, the site source in popsim, and there’s an arXiv draft in the works. Beat 82.0% and open a PR — I want to see it.