00:02:37 · Joon Sung Park, Molly O'Shea
[00:02:37] S1: Right, and tell me a little more about what makes up the simulation. What do you mean? [00:02:42] S0: Like, could humans be living in a simulation? Just like we all have the same, we're all programmed the same way, we all have the same psychological behaviors, like it's just everything is kind of already written, like what does it mean? [00:02:55] S1: So the way I often think about a society or an experience like what we are having right now is there's always this quantum building blocks of people's tastes and behaviors. So it might actually be things like my inherent characteristics. People's risk tolerance, not really known to change. There are certain aesthetics that I really enjoy. And these all come together and rebuild individual building blocks of these larger societies, which is individuals. So really, when we think about simulations, we think about individuals first. Can we create simulations of individuals that faithfully and accurately represent a diverse persona and characteristics and behaviors of that person? Then to create a civilization wise simulation, that basically means we put all these people into a same environment, let them interact with each other. And the part that's always complex and chaotic is what actually gets born in their interactions. So that's ultimately, I think, what building simulation here in terms of cytoscalability is going to look like. [00:04:01] S0: And so this all started with Smallville? It did. Okay, explain what that is and like your whole background behind that.
00:04:10 · Joon Sung Park
[00:04:10] S1: For sure. So Smallville was an experiment that we run back in 2023 and 2022, where we had the initial observation that these large language models, initially we were using GPT-three and then GPT-3.5, the models were fairly janky. But you could see a glimpse of opportunity where these models were trained on such broad human data that when you poke at the right angle, you can actually extract a lot of human data that appeared in its training data. This might be from the web, Wikipedia, some of the social media. And what we really start to ask ourselves is, can we actually use this to create an interaction paradigm that is truly new, that goes beyond just simply generating context or classifying different tasks, but can we actually generate truly new behaviors? And we decided that the most ambitious version of this is actually to create a simulation of an entire person. And then once we had an entire person simulated, can we actually go create an entire community? So Smallville is an experiment that came out of that effort. So it was a game town. Computer scientists love games because it's an amazing way to demonstrate new capabilities. So we created a game town that looks a lot like what Pokemon, Sims, like these kind of towns might look like. And then we populated this game town with 25 agents. And these agents all had different names, different characteristics, different life stories. And they were basically language models paired with what we called back then an agentic architecture that had memory, planning, and reflection. So they would actually wake up in the morning, do their routines, go to work, have relationships. The simulation itself actually takes place the day before Valentine's Day. So these agents will actually come together to have party. So there is a character that's actually organizing Valentine's Day party. There are characters that are asking themselves out. And none of this was preprogrammed, but what we could see is that we can basically create simulations of these people with emergent phenomena just coming out of the simulation. So that was the experiment that we ran in 2023. That was Smallville. That particular work ended up introducing a few key concepts. One was that we can create simulations of people, and the other was the core agentic architecture and that we can create agents using generative AI models.
00:06:29 · Joon Sung Park, Molly O'Shea
And it was very early days of these agentic workflows, So we could really start to see them becoming possible. And then we've been building on the technology basically for us to then create simulations that are more and more accurate and that's more diverse and that's bigger in scale. [00:06:46] S0: How do you test the validity of this? Like, do you know if it's actually working?
00:06:51 · Joon Sung Park
[00:06:51] S1: So our ability to test actually has shifted quite a bit also alongside the technological progress. On day one, when we created Smallville, what we actually did was we interviewed them. We interviewed the agents. What we've done is we pulled individual agents out of this game town, set them in this hypothetical chair on a command line, and we would basically ask them questions. And these questions would probe certain qualities of these agents that's similar to how people would reason about different tasks in their lives. So simple question, tell me your name. And then we would ask, tell us your life story. We would then ask questions like, who's the one person you want to spend time with? And then we would then ask, if you had to get that person a present, what would you get them? We would ask these questions because these are questions that when you think about human brain and human mind, we can only answer these questions because we can remember, because we can plan ahead, but also because we can reflect. Then what we ended up doing was we generated all these responses, and then we actually had human evaluators come in and take a look at these responses and said, given the experience that the agents had in their game town experience, are these experiences or insights actually human like? So we would actually basically have human evaluators compare between human authors' generation and these agents' responses to see which one's more realistic. And we've actually done quite well that these agents were actually quite realistic. And the way to sometimes tell is, imagine you're asking this agent, who do you want to hang out with the most? It's a little bit strange if the agent says whoever they happen to see the most on the street. The agent should likely say the person they actually care about. Who do you have special relationship with? What kind of conversations did you have with people? What actually stood out? And that agent should really reflect on that and pick those people. So those are some of the experiments that we ran, and those are the signals that we were able to get. But then over time, we wanted to go beyond that. So here, really the initial evaluation really was about believability that we start to see what we call the illusion of life in these agents.
00:09:08 · Joon Sung Park
But as we mature as a field, what we started to do was to run evaluation that is now about more of a faithfulness and accuracy. So the experiment that we ended up running following that was we ran an experiment where we then took this architecture to create simulations of thousands of real people. So that started with us actually going out, recruiting a large number of real people that sample representatively from The US. So it's a thousand representative sample of The US population across gender, race, ethnicity, and so forth. And then we actually collected data from these people. So we would actually do this time human interview. So we actually built an AI agent that can talk to voice to voice with people, have those agents talk to people voice to voice. We would ask questions that's a little bit like the questions we ask agents in Smallville. Tell me the story of your life. Where did you grow up? What were some of the hard decisions you had to make? Who matters to you? We would use that data to then train AI agents that can basically be the twins of these people. And then we would send people away for two weeks, bring them back to a lab. And then this time we would actually ask people to complete a battery of surveys, behavioral studies, respond to stimuli, and had their agents predict what they would say. And at the end of 2024, for the first time really in the field, we could actually show that we can predict people's behavior with 85% plus accuracy with these agents. That's ultimately what gave the company foundation for today's technology. Of course, the technology itself has shifted quite a bit, but that was the initial building block that we could see that, ah, models and these agents can now actually predict human behavior, and this is an amazing building blocks. But at the same time, it's large data operation that we do not simply create simulations out of vacuum. We believe strongly that the simulation has to be grounded, and it has to be evaluated alongside people. So it really is both.
00:11:09 · Molly O'Shea, Joon Sung Park
[00:11:09] S0: I have so many questions. Do you sometimes inject like anomalies or things that could create some factors that are not predictable? And then my second question, which I wanted to be my first question, is human behavior predictable? [00:11:28] S1: In the training data, we actually do inject a lot of behavioral knowledge. It's not just about what they say, but it's also really about what they do. And there's a lot of irrational behaviors in human tendencies. So a lot of things that we study in economic behavior game, randomized control trials, they all try to cover those. At the same time, there are certain things in people that's really not known to change. Your risk tolerance, as I briefly mentioned, it's not really known to change in a person. So once we are able to learn that, then this actually becomes a trait that becomes very predictable. So it really is about understanding the gap between what people think they are and what they actually do in their life. And really the modeling challenge here actually is really learning that translational layer. How do you take observed behaviors or attitudes into their fundamental tendencies? [00:12:27] S0: And what is in that gap? Is it time? Is it events around you? Is it like, how do you determine that gap that ultimately creates the outcome?