Simulating Humans at Scale: Simile's Joon Sung Park

**Joon Sung Park** (0:00)
I am somebody who is quite inspired by science fiction. And when you read science fiction, that covers societies that have progressed far enough in its technological maturity, you always see two pillars.
You have some version of AGI, and you have some version of simulations that really help guide the society. I do see an opportunity today to really take the first crack at building the simulation. I would not have said that even five years ago, but that is a conviction that we have built up over the years as we are going deep into this research.

**Sonya Huang** (0:52)
Today, we're delighted to have Joon, founder and CEO of Simile.
Simile is building an applied AI lab, simulating human behavior and societies. And I'm very excited to have you here to discuss what you're building.

**Joon Sung Park** (1:04)
Same here, thank you for having me.

**Sonya Huang** (1:05)
Okay, take me back to April, 2023, Stanford, California, specifically Smallville, Stanford, California. What was that?

**Joon Sung Park** (1:13)
So Smallville was a project that we were running at Stanford, where the idea was that we made this observation that large-lingual models can now encode a lot of human behavior that is imbedded in its training data, from the web and social media and so forth, that if you sort of probe at the right angle, you can actually get a lot of microbehaviors out of these models.
So given a very specific demonstration or description of a situation, what would person X do? And it would actually generate really interesting behaviors. We found that to be so interesting, and we found that to be the ingredient that we had been waiting for, for creating really complex adjunctive behaviors. So Smallville actually was an experiment where we decided that if we push this as far as possible, what would a society that is created by these agents look like? So we basically created generative agents that is paired with generative AI model, with memory, planning, and reflection, to basically create this lived experience of agents living in the small town. So Smallville was basically a game town of 25 agents living in it.
Individual agents had a description of persona, but they would actually wake up in the morning, do their routines, go to work, actually have relationship, sort of like people would, and they would actually have emergent phenomena, like having parties and so forth. So that was the experiment that we ran.

**Sonya Huang** (2:29)
What was the most surprising things to come out of the experiment?

**Joon Sung Park** (2:34)
So one of the surprising things was, so the experiment, the simulation itself, actually sets place the day before a Valentine's Day.
So you actually see these agents, one of the agents actually thinking, well, I run a cafe, so she's a cafe owner, her name's Isabella. She goes and thinks, it would be great if I can do a Valentine's Day party, where we invite a lot of friends, customers. So you actually see her on the day before Valentine's Day, going around, actually gathering materials for the party, actually telling our customers, hey, we're going to have this party, please come. And on the day of Valentine's, you actually see this immersion party that actually gets formed, with all these agents coming to the, to the basement cafe.

**Sonya Huang** (3:13)
Did anyone not get invited?

**Joon Sung Park** (3:15)
Well, some of the people did get the invitation, but they forgot. That's one thing that did happen.
Some of the agents did not explicitly get invited, but we had one agent who got the invite, Klaus, who decided to ask his crush out on a date. So he would actually bring in the date, they would actually have a party at this cafe. So quite surreal.

**Sonya Huang** (3:33)
So how did you end up building Smallville in the first place? Like, were you studying kind of human psychology and social behavior, or was this coming from, was this coming from the kind of customer back, or was it coming from the technology out?

**Joon Sung Park** (3:45)
So my particular team has been excited about simulations, and we saw the visual simulation failure early on. So my career as a researcher at Stanford really started back in 2020 That was the year when GPT-3 was about to come out. It wasn't quite there yet, but it was just about to come out. We started to get its first demos. And my first year, we wrote this paper called Opportunities and Disks of Foundation Model, alongside many of the Stanford researchers, and it was led by one of my co-founders, Percy Liang, who is now the head of the Center for Foundation Model at Stanford. And when we were writing that, the part that I was really focused on was, well, here's a new class of models that we have not seen in the past, that these models that can be very generalizable in ways we didn't quite have in the past. And I got into thinking, well, if we can imagine the kind of interaction we can create with these models, what would that be? And many of my colleagues back then were surprised that these agents or these models can do classification or a simple generation. And that was really incredible to see because these models didn't really know or wasn't really taught to do that. But the part that was surprising to me wasn't that these models can do that, because from an interaction perspective, we've known how to do this for a long time. The interesting part was, well, these models can actually encode human behavior. What does that mean if we were to push this as far as possible?