Latent Space

Joon Sung Park on simulating imperfect people to shape the future

Joon Sung Park· Co-founder & CEO at Simile
·~71 min·English·Latent Space
AgentsReasoningAI Company
TL;DR

Stanford's generative-agents author Joon Sung Park explains why his company Simile models real, imperfect people rather than idealized reasoners — from a validated 1,000-person study toward the vision of simulating all 8 billion of us.

01Origin Bet

The Time Machine Game

Before building the assistant everyone else chased, Park's team made the opposite bet: you have to model real people first.

what if we can just recreate the world that we live in? I mean, it's really hard to get more ambitious than that.

— Joon Sung Park, Latent Space
Key Insight
In 2023 everyone raced to build the agent; Park inverted the dependency and treated the user model as the prerequisite, not a feature. His own throwaway example carries the whole thesis: an assistant that orders pineapple pizza for someone who hates it hasn't failed at ordering — it has failed to model the person.
02Core Distinction

Simulation Is Not Prediction

Simulation is not prediction: decision-makers want the sequence of steps that changes an outcome, not the predicted endpoint.

what is the path that we need to take now to get to that particular future and that's what simulation allows you to do

— Joon Sung Park, Latent Space
Key Insight
This quietly redraws the product category. A predictor competes with forecasting tools; a 'path-to-a-goal' engine competes with strategy consultants — a far larger, stickier position, and one an LLM-as-oracle cannot occupy because a single answer is not a plan.
03The Data

The Three Data Buckets

A behavior model needs three kinds of data, and the most valuable kind — the cause behind a choice — is the hardest to get.

Because the world is our ground truth, but it happens once.

— Joon Sung Park, Latent Space
Key Insight
The scarcity is the moat. Interviews have to be collected and observations can come from transactions or the web, but cause-and-mechanism data barely exists — reality runs each experiment exactly once and never with a clean control. Whoever manufactures that data at scale owns something no web crawl can reproduce.
04Method

Stakes Make It Behavioral

The line between what people say and what they do comes down to one thing: whether the stakes in the decision are real.

if the stake in your decision is real. That's ultimately what makes it behavioral.

— Joon Sung Park, Latent Space
Key Insight
This is why a survey and a simulation are not the same product. Attitudes are cheap and unreliable; behavior under real stakes is expensive and load-bearing. It also explains the lab spend — actually delivering the item a participant 'buys' isn't running a focus group, it's buying ground truth.
05Training Objective

Models as Dumb as I Am

Frontier models are trained to be flawless reasoners; Simile trains for the opposite — models that make the same mistakes people do.

what we're trying to create are models that are as dumb as I am

— Joon Sung Park, Latent Space
Key Insight
This is the part most people miss: alignment work pushes models toward the ideal reasoner, which makes them worse human simulators. Simile is optimizing against the grain of the frontier — a different objective on different data — which is exactly why a bigger base model won't automatically win this problem.
06The Evidence

85% as Accurate as You

Simile's digital twins reproduce a person's answers 85% as accurately as that person reproduces their own — where frontier models fall to 20–60%.

we basically could replicate people's behaviors and attitudes 85% as accurately as people would replicate their own

— Joon Sung Park, Latent Space
Key Insight
The subtlety is the denominator: the 85% is relative to how consistently people reproduce their own answers, not 85% raw agreement — so the real ceiling is human variability, not perfection. That reframes 'how good is the twin' as 'how close to the natural limit,' a bar frontier models don't clear on the niche populations customers actually care about.
07Why Bespoke

You Can't Shortcut People

You can't shortcut human data with a demographic dot-product — that only retrieves priors the model already had.

You're at that point merely retrieving the knowledge that is already embedded in the model in the model parameters

— Joon Sung Park, Latent Space
Key Insight
If the demographic dot-product worked, simulation would already be solved — and Park concedes exactly that. The bet is that the decision-relevant behavior lives in the long tail the priors miss, which is precisely what can't be retrieved and must be collected. When it is collected well, he says an early scaling law appears: more human data and compute buy predictable gains.
08The Vision

Simulate 8 Billion People

The end state isn't market research; it's a simulation of 8 billion people to steer society's hardest decisions.

Can we create a simulation of 8 billion people living on earth?

— Joon Sung Park, Latent Space
Key Insight
Calling simulation AGI's 'twin pillar' is a positioning claim as much as a technical one: it moves the company off the $100B market-research line item and onto an open-ended 'inform every decision made about humans' TAM — and it justifies, in his own framing, spending foundation-model-scale compute on a single societal run if it could answer a problem like climate change.