模拟：新的扩展定律——Joon Sung Park，Simile AI

When we first dicsussed the Summer of Simulative AI in 2024 we knew it would be a brief summer, but it has recently come back with a vengeance with SimGym in April and now Simile AI’s $2B Series B, backed by GreenOaks and Index Ventures with prominent backers like Fei-Fei Li and Andrej Karpathy, running tens of millions of simulations for Fortune 100 clients like CVS and 85–99% accuracy vs human focus groups.

Time to catch up on why this Second Summer of simulation is working!

From creating Smallville, the landmark 2023 paper on Generative Agents that showed AI characters could remember, plan, socialize, and develop emergent behaviors, to now building foundation models of human behavior, Joon Sung Park is trying to answer a much bigger question: what if we could simulate the world before making decisions in it? In this episode, the Simile co-founder and CEO joins us to unpack the path from generative agents to digital twins, why today’s frontier models still fail to capture how humans actually behave, and what it would take to eventually simulate all 8 billion people on Earth.

We go deep on Simile’s approach to modeling human behavior: long-form interviews, observational and transaction data, randomized controlled trials, population-level and individual-level models, and post-training on the causal mechanisms behind why people make decisions. Joon explains how his research created digital twins that reproduced human behavior and attitudes 85% as accurately as people reproduced their own responses, why models optimized to be rational can be bad simulations of irrational humans, and why understanding “social physics” may require changing model weights rather than simply prompting frontier LLMs.

We also explore the much larger ambition behind simulation: testing products and policies before deploying them, finding counterintuitive paths toward desired outcomes, modeling emergent behavior across entire societies, and potentially tackling problems like climate change, democratic instability, and UBI. Joon reflects on scaling laws for simulation, the economics of data-center-scale simulated worlds, the connection to Thomas Schelling and psychohistory, why simulation is surprisingly similar to painting, and whether we might already be living in one.

We discuss:

- How Smallville and Generative Agents led to Simile

- Why Joon’s team asked: “What if we can just recreate the world that we live in?”

- Why useful personal agents require deep models of their users

- Memory architectures, Markdown files, and the limits of prompting

- “Social physics” and behavioral foundation models

- Why web data captures what people say more than what they actually do

- Interviews, transactions, observational data, and randomized controlled trials

- Why predicting the future matters less than understanding how to shape it

- How Simile creates representative simulated populations

- Simulation versus prediction and the connection to Foundation’s psychohistory

- How to evaluate simulations instead of simply stacking LLM hallucinations

- Creating digital twins of 1,000 real people and reaching 85% behavioral accuracy

- Why frontier models can struggle to reproduce real human behavior

- Why good simulations need to reproduce human biases and mistakes

- Post-training models on randomized controlled trials

- Population-level versus individual-level simulation

- Scaling laws for human simulation

- The long-term ambition to simulate all 8 billion people on Earth

- Whether simulations could help solve climate change or detect collapsing democracy

- Thomas Schelling and the history of agent-based modeling

- Why future simulations could require an entire data center

- Multi-agent simulations and what happens when simulated people interact

- Replacing expensive human panels with synthetic populations

- Why market research is only the starting point for simulation

- Why Joon sees simulation as surprisingly similar to painting

- Using simulation to study questions like UBI

- Whether we are already living in a simulation

- Why AGI and simulation may be the twin technologies of advanced civilizations

Joon Sung Park

- LinkedIn: https://www.linkedin.com/in/joonspark

- X: https://x.com/joon_s_pk

- Website: https://www.joonsungpark.com

- Simile: https://www.simile.com

Timestamps

00:00:00 Introduction and Joon’s Path from Art to AI

00:01:46 Smallville, Generative Agents, and the Origins of Simulation

00:05:03 “Let’s Just Create a World” and the Future of Personal Agents

00:09:53 Social Physics and Behavioral Foundation Models

00:14:08 Prediction vs. Simulation: How Do You Shape the Future?

00:16:59 How Simile Models Real People and Populations

00:25:35 Evaluating Simulations, Digital Twins, and 85% Accuracy

00:30:23 Post-Training Models to Reproduce Human Behavior

00:40:04 Scaling Laws and Simulating 8 Billion People

00:43:10 From Schelling to Society-Scale Agent Simulations

00:46:13 The Cost and Economics of Simulating the World

00:52:05 Real-World Use Cases, Synthetic Populations, and the Market

00:57:27 The Future of Simulation, Painting, and UBI

01:04:23 Are We Already Living in a Simulation?

01:06:08 Building Simile and Hiring

Transcript

Introduction: Joon Sung Park, Simile, and the Story So Far

Vibhu [00:00:00]: Today, we have Joon in the podcast. Excited to kick this one off. Very exciting company. I wanna kick off and ask you the question, talk us through the story of your life. How have you gotten here?

Joon [00:00:13]: Yeah, for sure. I’m really excited to be here. A story of my life. So I was born in Korea, and I lived there for a good 11 years or so of my life, and then my family moved to Boston. So we moved when I was 11, and my parents were doctors, so they were going through their postdoctoral studies. My dad was a surgeon, so he was doing his sabbatical years at the Boston Children’s Hospital. So I grew up there, not too close to tech. I was very much a music and artsy, painting kind of guy.

Vibhu [00:00:49]: Painting.

Joon [00:00:49]: Exactly. I got into painting a little bit later, in high school, but that’s what I used to do. And then I grew up mostly in the East Coast after Korea. So I lived a good number of years in New Hampshire, and then I went to college in Pennsylvania. And I got into more of this tech scene, in college. So I was originally trained to be an artist. I thought that would be my professional career. So it wasn’t a hobby. It was like, “Hey, let’s make a living out of this.” And then gradually, I got really interested in this idea of, hey, the greatest artist often creates their own medium, and the best medium that we had available today was in computation. So I decided to go deeper into that, and one thing led to another, and we can go deeper into this, but I decided that research was something that I gradually got interested in, and here I am.

Smallville, Generative Agents, and the 2023 Breakout Paper

Swyx [00:01:46]: So there’s a lot that you packed into the research components. You had one of the best papers of 2023, which was the generative agents paper, commonly known as the Smallville paper.

Swyx [00:01:58]: Feel free to call back to anything else that you mentioned, but most people would have heard of you from this. Do you have any statistics on how many people have, like, read it? arXiv gives you something, right? Some stats.

Joon [00:02:10]: Yeah, it’s a good question. How many people have read it, I’m not sure.

Joon [00:02:14]: I know we do keep track of citations, and they are going up quite fast.

Swyx [00:02:23]: Yeah, Google Scholar has 7,200 citations.

Vibhu [00:02:25]: I feel like it made a bigger hit than that, and it was a pretty instrumental paper. It got cited so many times.

Swyx [00:02:34]: It is frequently the answer when people ask, “What is the best paper you’ve read recently?” It’s this one.

Vibhu [00:02:39]: I thought the memory component was pretty underrated. It was a very good early memory system, and one of the biggest papers.

Foundation Models and the Search for Killer Applications

Joon [00:02:47]: Yeah, so maybe I can talk a little bit about how this particular paper came together. So when I got into research, it was back in 2020 when I started my PhD program at Stanford, and that was the year, when we were about to get GPT-3 to be available. So we already had GPT-2, and you could sense that there was this new class of models that was just becoming available in the market, and the team got very intrigued. And the general consensus was, “Well, is this model going to be useful for anything?” “It’s really strange that these models are not trained to do any particular task.” But we decided to take a bet. So a large group of scholars at Stanford, and it was led by one of my co-founders, Percy Liang, and we came together

Swyx [00:03:35]: Who coined foundation models.

Joon [00:03:36]: Who coined the term foundation models. We wrote this paper, where that term came from called Opportunities and Risks of Foundation Models. And during that process, really the thing that I started to think deeply about was, here is a model that is fundamentally new in our ecosystem. The reason why this was new was it wasn’t, again, trained to do anything in particular, but its premise was it could do anything and everything. It was like a stem cell, if you were to take a biology analogy. And I got really interested in this idea that, well, if we were to really think about what are the killer applications that this particular technology would enable, what would that be? Many of my colleagues were using this for simple classification, simple generations. Interesting that these models can do that, but from an interaction perspective, not that interesting. We’ve known how to do that for many decades. And what we came down to was these models are trained on this very broad data from the web, right? So these are human behavioral data. It’s social media, Wikipedia, all these data. So if you poke at the right angle, then you could see human behavior that would just pop out that’s quite realistic, and we’ve never seen that before.

The Time Machine Game and Recreating the World

Joon [00:04:45]: So that got us really interested. The exercise that we decided to do, with this particular group of colleagues, Michael Bernstein, Percy Liang, and myself, who ended up becoming my co-founder at Simile, we sat down and we played this game that we call the time machine game.

Joon [00:05:03]: Imagine we were to get on a time machine and fast-forward 10 years and look back. What would have been the single application that will have mattered that would be the most interesting and inspiring? And when we thought, “Well, what if we can just recreate the world that we live in?” it’s really hard to get more ambitious than that. Like, let’s just create a world.

Joon [00:05:24]: And that’s where we started. And initially, we had this paper that was a precursor to the generative agents paper called Social Simulacra.

Swyx [00:05:32]: Before you go further, were there other candidates for the most ambitious thing in the time machine exercise? What was number two or number three?

Personal Agents, User Models, and Why Simulation Came First

Joon [00:05:44]: There is a close second that we were considering, which ended up becoming more of these automation tools, especially the vision around really personalized agents that would do things for you.

Swyx [00:05:59]: That’s also happening.

Joon [00:06:00]: It’s also happening. But it was interesting for us, right, in that the reason why, we decided to go with the idea of simulation, one, I was a huge science fiction nerd, and this idea of creating simulation, I was personally really just fascinated. I loved the idea. It’s really cool to see, like, a game town like this and just see these agents live in it. But at the same time, my bet was if you were to create a really amazing personal assistant out of this technology, what you need first is an amazing model of your users. So I told a model, “Hey, can you go buy late dinner for me?” And it orders Hawaiian pizza, and I do not like pineapples on my pizza. Then it totally failed. The way for it to not make that mistake is only by having a deep understanding of who I am. And I gave a very simple and dumb example here, but you can imagine how this core understanding of people is instrumental. This is how, if we have our family and closest friends, they have a good mental model of who we are. That’s the basis of our social connection. So our bet also was this technology around simulation, creating accurate representation of people ought to precede the more complex agents that would automate the world that we live in. So that was the bet. But that was a very close second, and I’m still very much fascinated by it. I think there’s a lot of interesting work that’s going around. My hot take here, though, is I don’t think we’ve seen a true personal assistant that’s useful, in ways that meet the ambition of that particular line of work. I think there are early applications that are interesting, and if you talk to even ChatGPT nowadays or Claude, they know a lot about us. So a lot of the generation it’s doing, I do think it’s much more tailored, but I think the ambition is quite large in that field, and I don’t think we quite have all the right ingredients just yet.

Swyx [00:08:01]: So OpenClaw and these personal agents, what do you want to see from them that they don’t currently have?

Memory, Markdown, and the Limits of Prompting

Joon [00:08:09]: I do think it’s slowly getting there, but I do generally want them to have much deeper understanding of the person. Right now, you look at the models. OpenClaw, what it’s leveraging is a Markdown file, and I think it’s quite clever, right? So if you look at the generative agents paper, this was the same intuition that we had, where initially when we were creating the memory architecture for the generative agents, and, like, this is, like, back in 2022, so we didn’t really quite have the idea of even agentive architecture or the term agent. But the intuition that we shared with some of the work that’s coming out today was we initially thought, “Well, do we want to make the memory into, let’s say, knowledge graph? Do we want to train a bespoke model?” All of these things. And what we decided to do was, “No. Just forget about all this.” These language models are quite good at modeling text and understanding and reasoning about text. So just put everything in a Markdown file or a text file. You’re done. I thought that was quite interesting that we could do that, and there’s a lot of strength in doing that. But also, there are limitations. It’s the way you retrieve and make sense of data that’s extremely large, it takes a lot of work. So I think that technology is getting better. I also do, however, think, there are certain things you just cannot shape just by prompting the model. So to some degree, you do need to touch the parameters of the model itself. So there is this work that I do think does need to happen, and it is happening. The question is, how far can we take it? How do we source data, and how do you also create an ecosystem where people are continuously feeding data to this model so it’s learning about you?

Vibhu [00:09:50]: What’s the intuition between why you need to do it in the model?

Social Physics and Behavior Foundation Models

Joon [00:09:53]: My intuition behind the actual when do you train or even post-train a model versus just prompt a model is if the model has to learn the underlying physics of the world that it’s operating in. So it has to learn new social physics. The places where it doesn’t have to train are the places where it already has the physics. We trust the physics. It already has the base statistics, but it’s just trying to react to an environment. Then I think you can just prompt your way into getting the actions out of it. I don’t think the models that are out in the open have yet learned the complete mapping of social physics of humanity. This is one of the core theses of Simile, right? And one of the core reasons why that is the case is if you look at the data that the model was trained on, these models were trained on the web data, like, whatever was available on the web. And these are really interesting data sets, but they are fundamentally the self-exposed attitudinal data with some behavior data that’s sprinkled around here and there. And it has yet to learn the really deep behavioral nature of people, not just what people say they do online, but what they do in real life. And this is one of what I would consider to be the dark knowledge of humanity that we haven’t quite captured. And it’s these data that would also need to get factored into the model creation.

Vibhu [00:11:21]: You call it behavior foundation model.

Vibhu [00:11:23]: There’s a good one-liner here, but outside of that, what type of data do you need? What are you changing on the model level? How do you go about modeling, doing a behavior foundation model?

The Three Data Buckets: Interviews, Behavior, and Causality

Joon [00:11:35]: We think about data in three buckets. So one bucket is interview data. It’s quite interesting. Rich qualitative data is interesting. It’s not behavioral, but we would literally ask people, “Hey, tell me the story of your life.”

Vibhu [00:11:53]: It’s just what we’re doing here exactly.

Joon [00:11:54]: The question that you all asked at the beginning of this interview literally is the question we also ask. And we ask our participants to go a little bit deeper, than how far I went. Maybe I can give more of my life story in lieu of this. But the reason why that data is interesting is by learning about this very long-tail information about people, you get a lot of texture around this model, like, this person as a model. So even understanding their childhood memory or even their trauma, their first love, these things, quite informative in ways that’s really hard to predict. So that’s one. Then there are two tranches of what I would consider to be the behavioral data. One kind of behavioral data is observational. So these might be like transaction data, or these might be data that you can get by scraping the web, right? So you can imagine why these data sets would be interesting, right, because they give you the base statistics of people’s behavior.

Joon [00:12:55]: But then there is the last category of data, that I personally think is perhaps the most important, which is the data that describes the causal mechanism, the whys of people. Some of this is covered by the interview data, the qualitative, because people talk about why they made certain decisions. But really, where you get to see the most behavioral aspect of this is in randomized controlled trials, like RCTs. Imagine you have the same setup, but you have a few different variables that you are trying to tweak. Can you get realistic human behavior out of it in ways where, imagine you had this particular option. Imagine you’re even trying to choose whether you’re going to drink coffee or not. The day you drink coffee versus the day you didn’t drink coffee, does your behavior change? That’s a data set that describes a causal mechanism. This is quite important in modeling people. The reason why this is important is oftentimes when people come to us, or not just to us, but the reason why people are interested in simulation isn’t because they want to predict the future. If you’re trying to win against the stock market, predicting the future is interesting.

Prediction vs. Simulation: Shaping the Future

Joon [00:14:08]: But most people, most decision-makers, what they want to know is, how can we shape the future? It doesn’t really help you to hear that your sales are going to tank in two quarters. They’re just gonna say, “Wow, that sucks.” What they want to know is, well, what do we need to do now to avoid that future? That’s the causal mechanism. And this is also very hard data to come by, right, because the world is our ground truth, but it happens once. So in a very controlled setup where everything is equal except for one variable, this kind of data set rarely happens. So this is a reason why this data set is both hard to come by and quite important if you’re trying to model human behavior.

Swyx [00:14:50]: So behavior, I think, is the hardest data set to acquire. What is out there? What is even possible? You’re not going to know a lot of details about my life. I don’t even have data for myself on my own health or habits, and I just don’t log everything. So how can you have that data?

Joon [00:15:14]: So we run a lot of randomized controlled trials.

Swyx [00:15:17]: But you put people in the lab, they watch them sleep, or what?

Joon [00:15:20]: We do care a lot about the consent process. People know that we invite them to be a member of this community to both share data and have themselves represented in different forms. But we bring a lot of people to the lab, or virtual lab, where we design experiments that would pose them real behavioral decisions. And often in these experimental setups, what makes the difference between what is attitudinal versus behavioral is whether the stake in your decision is real. That’s ultimately what makes it behavioral. So in these setups, we are inspired by our colleagues in social sciences, psychology, and so forth. So when they run studies, the techniques they utilize is imagine there’s an online store that you’re inviting people to come by. Then whatever they purchase in this experiment, they actually get that item delivered. Like, these are the things that make the stakes real. So we run a lot of these experiments, and we also do partner with firms. Right now, we also have customers who are quite excited to at least give us a glimpse of the behaviors that their users exhibit so that we can get a little bit deeper understanding of how people behave in these different platforms.

How Customers Use Simile: Populations, Queries, and Experiments

Vibhu [00:16:39]: I think on the customer side, they have a lot of data about their users, who has bought. They have the action data.

Vibhu [00:16:47]: Can you walk us through an example of what someone comes to you for? What questions would they want solved? Do you customize a model for them? Do you have something off the shelf? What does that look like?

Joon [00:16:59]: Today, when people leverage our models, it’s often to better understand the population of their interest. So usually, the start of the relationship, we come together and hear about what population they want us to model, right? So it might be that if you’re a CPG company that’s selling to all of the US, then maybe it’s fairly straightforward. You want to model the gen pop of the US. But at the same time, if there is a vertical or if there’s a market that they’re trying to go into, imagine, they want to better understand, let’s say, people in their 20s and 30s living in California. That’s a much more specific population. So we hear about this population, and we go recruit these people, with consent, and with incentives, and we collect some of their data and create a model of these people. Then what our product allows you to do is query them. So it can take as input a filter that is a description of the population that you want to talk to, just like the one I just mentioned, and an environment. The environment can literally be survey questions, behavioral experiments, It can be A/B testing. Oftentimes, the core use cases are things like concept testing, to start with. But also, people sometimes want to do focus groups or one of the fun use cases that we also serve is even modeling things like earnings calls for public companies.

Joon [00:18:21]: So these are the use cases that we often start with.

Swyx [00:18:23]: Concept testing, is that an established term? I’ve never heard of concept testing.

Concept Testing, Gallup, and Politics

Joon [00:18:27]: Yeah. So it has to do with they have, let’s say, different messaging, different products, different ideas.

Swyx [00:18:32]: It’s like a marketing exercise.

Swyx [00:18:33]: Okay, got it. Got it. Politics?

Joon [00:18:36]: We do, have a strategic partnership with Gallup, and of course, Gallup is deep into policy space and so forth. Right now, we have not worked deeply with politics, like that area just yet, however.

Swyx [00:18:49]: I’m curious if there is demand or if they really would have different needs that somehow fundamentally don’t mix with your existing, users or people.

Joon [00:19:00]: I think there’s certainly demand.

Joon [00:19:02]: But we are very much mindful of how this technology gets adopted and the societal impact that we’ll end up having with this technology. And I do see politics as an area where a company has to be particularly thoughtful about the way they operate and make impact. So this is where we also want to make sure that we form enough of guardrail and perspective on how to leverage this technology before we go on to serve markets like the politics.

Swyx [00:19:29]: I’ll give people an example. one of my favorite shows is The West Wing. I don’t know if people have watched.

Swyx [00:19:34]: One of the key storylines is, like, the president has, multiple sclerosis, but they haven’t. they need to figure out how to disclose it. So they run a poll with a fake governor and ask people to respond on the poll,

Counterfactuals, Polling, and When Simulation Is Useful

Swyx [00:19:47]: They try to make decisions based on the results of that poll on, like, how well they’ll be received, like where, how should we play this?

Swyx [00:19:54]: And I’m like, well, I think those counterfactual things, I would use a simulation for this if I could trust it.

Joon [00:20:01]: For sure.

Joon [00:20:02]: In that show, how’d it go?

Swyx [00:20:04]: In that show, it was, like a foregone conclusion. They were like, “We know it’s bad. We just don’t know how bad.” And then the poll came back. It was like, “It’s really bad.” And then they just did it anyway.

Joon [00:20:14]: Part of it is to show, right? So you’re, you’re looking at the idea

Swyx [00:20:17]: Maximizing drama.

Joon [00:20:18]: How bad could it be? Oh, it’s horrible.

Swyx [00:20:20]: And to some extent, I think that is part of the trick of the, or the challenge or with being a customer of yours, which is that if I know it’s. if I roughly know and can intuit

Swyx [00:20:35]: What the effect is going to be, do I need you? What sensitivity of it, of effect do I need in order to make a decision, right? So for example, if I, my approval rating is 50%

Swyx [00:20:48]: And I, they have this negative piece, news item comes out, and it drops to 30.

Swyx [00:20:52]: If it drops to 20, if it drops to 40, do I care? No. It, I know it drops. It’s negative. So when do I care about simulations?

Joon [00:21:01]: You do something that’s clearly bad, that’s not popular, and people don’t like you, like, yeah, it’s like

Swyx [00:21:05]: You don’t need a simulation.

Joon [00:21:07]: Yeah. Well, so there are a couple of things. one is, there are use cases where, like every day, developers, designers, policymakers, marketers, every single day, they create assets. They create new products. And turns out, it’s many of the decisions in hindsight is obvious. Yes, of course this is bad, but we still run those studies because understanding the magnitude and understanding how acute something is quite difficult, even if, we feel like, of course, like this makes sense. this is the reason why we make so many mistakes. Like, every time somebody goes online and say something that has huge backlash, you look at that and like, “What an idiot.” However, it’s tough. That’s one. There’s also another aspect here, which is, again, this is the reason why simulation is different from prediction. In simulation, in the ideal case scenario. So what simulation is trying to show is it’s trying to show each step of the way or each step that we need to take to get to a certain outcome, right? So in the most advanced simulations, sometimes the next step that we’re suggesting might be quite counterintuitive. The analogy that I sometimes give, and I ground it in a more realistic example, but, I, as I mentioned, I’m a huge fan of science fiction, and I don’t know how, many of the audience members have read, like, things like the Foundation series by Asimov.

Simulation as a Path, Not Just a Prediction

Swyx [00:22:37]: Oh, yeah. We’ve mentioned psychohistory a number of times.

Joon [00:22:39]: Okay, fantastic. So I might be, talking to the right crew. If you read Foundation series, literally the first act is there’s a group of scientists who have found out that, “Oh, our galactic empire is going to collapse, and we’re going to have 30,000 years of unrest.” And they run psychohistory, the simulator that tries to teach them, “Okay, how can we keep this unrest to a 1,000 years?” And they plan this out, and the first step of that plan is to get the scientists who say, “Okay, this is coming,” exiled into this random place in this, galax- galaxy.

Swyx [00:23:18]: Terminus.

Joon [00:23:19]: Exactly. And that’s so counterintuitive. Like, what a strange move that you literally sent the group of scientists who was raising voice around this potential collapse of galactic empire into nowhere. How is that the right first move? Well, it turns out in this particular simulation, that was the move.

Joon [00:23:40]: It’s these things, right? And the reason why these reasoning is possible is because you’re showing the step function or each step that results in a particular outcome. So really what simulation allows you to do in its highest form is you give it not a problem or question, like what would people answer to the survey? That’s not what we do. What we tell it is, “Here is a goal that we have. In the context of foundation, we want to keep the unrest to a 1,000 years. 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. Now, translating that into real market, imagine you’re a automobile company and you’re about to release a, EV, and you’re trying to understand, well, how do we market EV, to make sure that our stock price goes up? But what if the answer comes down that, well, you can market your EV in XYZ way, but that might change people’s perception around the cars that’s not EV and make your overall sales to go down. Not very intuitive, especially all you’re trying to optimize is EV salesss, and that’s the only thing that you’re tracking, then that might result in a completely wrong solution, or at least different solution than what you would have expected, whether it’s right or wrong.

Joon [00:24:57]: That’s the power of simulation.

Swyx [00:24:58]: For listeners, we covered a similar topic with Mikhail Parakhin from Shopify, where they are working on SimGym. I don’t know if he ever talked to you about it. it’s very similar.

Joon [00:25:07]: I

Swyx [00:25:07]: The goal is increased conversion, but then the journey is very unusual.

Joon [00:25:12]: Journey is unusual.

Swyx [00:25:12]: Yeah. The-- He’s trying to look for interventions on a shopping trajectory, which is similar to what you’re saying. Like, it’s not about the attitudinal, is your word for it.

Swyx [00:25:24]: It’s about behavior.

Joon [00:25:25]: It’s about behavior.

Swyx [00:25:25]: And that’s exactly the difference, right? It’s, like, not about the near-term direction about-- but it’s more about, like, how do you affect multiple turns of interactions.

Vibhu [00:25:35]: You had a good quote at the start about this as well. It’s not about people wanting to know the outcome. It’s about how they can change it, change the way to get there, something like that. But I wanna take it back to how do we know this is grounded? Like

Grounding and Evaluating Digital Twins

Vibhu [00:25:47]: How do you run evals? How do you test that simulations come through? if I was to do the same thing that you described with, say, your favorite LLM, Opus, GPT-5.6, have some agent to map out these things

Vibhu [00:26:02]: How different are the answers we would get if I give it the same goal, the same objective, make a decent system? You’re saying that you need to change the model weight. You have your own solution to this. But how far off are we, and how do you check if it’s grounded? you have some interesting stuff on your site that points to how you run real evals, but if you could take us through that side. I think that’s one of the big concerns that people have. They’re like, “LLMs hallucinate.”

Vibhu [00:26:27]: “You’re just hallucinating layer after layer,” right?

Joon [00:26:30]: The way we do this, and this is the paper that we worked on after the generative agents paper that really became the, at least for Simile and also the field of simulation and synthetic panels, really became the foundation. Yeah, this is the paper. the paper is called Generative Agent Simulations of 1000 People. Here’s what we’ve done. For this paper, we brought 1,000 people that’s representatively sampled from the US to a virtual lab. And what we have done was we spent two hours collecting fairly wide-ranging data. In this particular study, we focused a lot on this interview data, that was, whose script was taken from this project called American Voices Project. And then we would also pair that with a lot of behavior data and so forth, whatever we can collect within two hours. And then we would send these people away for a couple of weeks. And during that time, I would use this data to create their digital twins. And I would bring the humans, participants back after 2 weeks and have them complete a battery of surveys, experiments, behavior studies. So we have the list here, which included things like behavioral economics games. We would run literally, like, Big Five personality test, General Social Survey. We would also go ahead and run the randomized controlled trials that were published on PNAS. And we would have their digital twins predict how the source individuals would have acted in these studies and surveys. And this is where we could replicate people’s behaviors and attitudes 85 percent as accurately as people would replicate their own. So that was the first really paper that gave this validated results that we can model individuals in an accurate way. And what we ended up finding now, of course, in AI space, so this paper came out at the end of 2024. AI space, a year and a half, 2 years, that’s a lifetime.

85% Accuracy and Why Frontier Models Miss Human Behavior

Swyx [00:28:24]: Yeah. Just, for listeners who are not seeing the YouTube, I just wanna say, like, the headline figure is 85 percent accuracy, like, which is a big improvement over all the other

Swyx [00:28:34]: Methods that you showed.

Joon [00:28:36]: But the part that was particularly striking to us, especially as we improved this technology even further, was the generative AI models like ChatGPT, Claude that’s coming out, it does give you the right foundation. However, what they do not consider is the true attitudinal and behavioral aspect of people, especially in the population that you care about. So what these models are really good at today is they’re trying to become the super rational, objective machines, right? So you go get their data from places like Mercor, Scale. You talk to professional programmers, scientists to create model that’s amazing at reasoning. That’s what they do. Simile doesn’t care about any of this. The models that we’re talking about here, what we’re trying to create are models that are as dumb as I am, right? So if I make some mistakes, the model has to make the same mistake.

Swyx [00:29:34]: Oh, that’s very hard.

Joon [00:29:35]: That’s very hard.

Swyx [00:29:36]: You’re solving Murphy’s paradox.

Joon [00:29:37]: That’s exactly. And this is a completely different data and training objective. This is also where we see quite a bit of discrepancy in the performance in human behavior prediction between the frontier models, Simile’s model, and the models being created in this space, where in some cases, the model performance of frontier models go all the way down to 20, 30 percent, especially if you go into that more niche population on topics that our customers would care about. On more gen pop, it might be around 50 to 60 percent. So it’s not very robust. Like, you wouldn’t want to make your decision off of these and these findings. If you can bring that up to 85 percent, that is ultimately what people end up getting very excited about.

Swyx [00:30:20]: Yeah. Do we wanna keep going on the paper, routes?

Joon [00:30:23]: Yeah, for sure. So the last one, was an interesting one. So this, paper was the follow-up paper that we had, to the 1000 agents paper, where the idea was now can we augment the models even further and post-train a model based on a lot of randomized controlled trials? So this was an interesting one. The data is always the most interesting part of modeling in many ways. The data that we got here was there’s this, there’s this platform called Open Science Framework. So some, the audience might be familiar with this. And there has been, especially in the social sciences over the past 5 years or so, there has been this concern around replicability of studies. And so it was a bit of a crisis, the scientists acknowledged, where we rerun the study and we don’t see the same finding.

Post-Training on RCTs and Replication Studies

Vibhu [00:31:12]: Oof.

Joon [00:31:12]: It’s tough. And the reason why it’s there-- that was often the case was there’s this survival bias where the papers that get published often need to maintain what we call the value of less than 0.05 in the experiments that we ran. That suggests that only-- there’s only 5% chance that the results that we saw is false positive. But the tricky part was all the papers that were not published, and there’s still a 5% chance that whatever we publish is totally just randomly generated. Like, there’s a 5% chance that, hey, this effect is not real, but it just happened to be real because of the sampling bias. So because of that, what scientists started to do was they started to register their studies. So before running an experiment, they would go to this platform and say, “Here is the data. Here is the population that we’re collecting, and here’s the hypotheses.” And they would just say, “Here is our hypothesis.” Like, “This is what we believe.” And you cannot retroactively change those hypotheses. This is what gives us more scientific statistical confidence that whatever effect that you ended up seeing is true. So that ended up creating this really interesting platform where there’s one platform that has now contains tens of thousands of real-world experiments and hypotheses. And a lot of these are really high-quality, like, professionally designed behavior studies and random- randomized controlled trials. So we got the data and the studies from this platform and used that to make a point. And this particular, model is not, something that we’re serving commercially because this was a part of the open science. But this particular data set, helped us make a point that by collecting a lot of these randomized controlled trials, that are really well-designed, we can make significant improvement in model’s capability to predict human behaviors. So that’s what this paper was about.

Vibhu [00:33:10]: Is this stuff done on a individual level? Like, do I need to tune the model per individual, per company? Is there foundation model changes and then some slight post-training? Anything you can share there?

Population-Level vs. Individual-Level Models

Joon [00:33:21]: So this particular model was trained. the data we had at the level of individuals, but this particular model was trained. We experimented with both. And this is what we end up doing at Simile too. We always train 2, distinct model. One is what we call the population-level model. The other is what we call the individual-level model. And both take very similar input, which is the description of a subpopulation or individual and a stimuli. In this particular work, we’ve done the same. Here, the results that we are reporting are much more geared towards individuals because we do think that is a harder task in many ways, but that’s what we have done.

Vibhu [00:34:02]: You seen anything on the questions that humans can solve that models can’t solve? So like

Human Biases, Mundane Choices, and What Models Miss

Vibhu [00:34:09]: Currently, it’s, I live 5 minutes walk away from a car wash. It’s a 10-minute drive. Should I walk or drive?

Joon [00:34:16]: Huh.

Vibhu [00:34:16]: The model will say, “Oh, walk to the car wash.” And, you don’t have your car.

Vibhu [00:34:20]: Is anything like this a problem in simulation? You would assume, like, very simple for human to think about, but if the model is saying you should walk to the car wash, anything here?

Joon [00:34:32]: It’s less, what can we solve, but I think it’s more about what biases or mistakes do people make that models miss. Like, imagine that you are, like the. When I was still at Stanford, I lived in Palo Alto. So it’s about, I would say, 40-minute walk from the campus. You ask the model, “Okay, let’s go home. What can I, what can I do?” It would likely call an Uber or, give me, the bus time. But for the longest time, I really liked walking back. And the reason why I wanted to do that was not for efficiency. It really helped me think. And I like to walk for, half an hour or 40 minutes or so a day, where I just get to, just think about ideas, research, just get lost in my thoughts. That’s very human activity. Unless the model has seen that and understands the importance of that activity, it would miss these kinds of features. So that I think, is fundamentally what we’re trying to model. Like, what is fundamentally human might not be the most efficient thing to do, might not be the right thing to do, but things that make us who we are.

Swyx [00:35:43]: I’m curious if, there are some data sets that you really want that would materially help you. One version of this may be interesting, which is more valuable to you to acquire as a data set, all of LinkedIn, all of Twitter, all of Facebook?

What Data Matters: Social Media, Transactions, and Facebook

Joon [00:35:57]: It’s a little bit hard to rank, in part because, there’s, there’s this product saying where no feedback is wrong because it teaches you something about your users. Doesn’t matter what feedback.

Joon [00:36:11]: I think it’s a little bit like that.

Swyx [00:36:12]: So just whatever is bigger.

Vibhu [00:36:13]: What about a different domain? Say it was. What about all of Amazon data?

Joon [00:36:17]: Oh, yeah.

Vibhu [00:36:18]: Shopping data, right?

Joon [00:36:18]: Shopping data. So Amazon data is interesting in that it’s very much behavioral, although, like, what people do on social media, you could squint and say that is also behavioral. But the transaction data is always interesting. It is also most commonly available, however.

Joon [00:36:33]: If we were to look at purely social media, like if you really, if I were, if I had to really pick, Facebook likely is interesting because I do think it is most a default version of people. Because you go to LinkedIn, it’s very much professional environment. So people put up their, they have their guards up, right? And that still is interesting because that is true human attitude and behavior, but it is not your base state. you go to Twitter- Twitter, people have their own crazy personas, or depending on who you are. Like, my Twitter profile and, persona is very much, initially was I was very much an academic. “Hey, I’m here to share my studies.” Now, I share, things that’s related to Simile. But Facebook is one of those more private space where people just connect with their friends. In that way, I do think it shows you a little bit more about who that person is. So if I had to pick, I’d likely pick, Facebook.

Swyx [00:37:30]: Yeah. And you’re interested in, like, the whole person and their background and philosophy. I, is it too clinical or too machine learning-oriented to just say this is just ways to inject variance and biases? The broad question, is, like, is this any better than a randomized, like, combinatorial explosion version? So we have a link to the Tencent

Billion Personas, Synthetic Demographics, and Bespoke Data

Swyx [00:37:54]: Billion persona paper, where they did not do any of the groundwork that you are doing.

Swyx [00:37:59]: They just did like a cross matrix of here’s all the professions in the world, here’s all the people, possible backgrounds in the world, do a dot product across all of them, and that’s it. That’s your prompt for a billion people.

Swyx [00:38:12]: This will do something. I don’t know if it’ll do what you do, but it gets you some way, some percent of the way there.

Joon [00:38:18]: So this was an interesting paper. Like, what I admired about this paper when it came out was the scale. And you do gradually want to be able to simulate really large societies and interactions. So the scale is definitely admirable. it is relying heavily on the known statistics that went into training the model. So to the extent that you believe that statistics is correct, this is not a bad way to go about this. But the thesis here, and this is something that we also have seen in the market, like if this works, then we have solved simulation.

Joon [00:38:54]: It,

Swyx [00:38:55]: Because I survey, like, okay, 5% of the US population is in construction.

Swyx [00:39:01]: The other 5% is in medicine, whatever, right? And then you just keep going down the list, and then you do the other side. 5% has, like, the big 5 personality

Swyx [00:39:08]: Of, like, neurotic or whatever. That’s it.

Joon [00:39:11]: That’s it. So if you believe that the underlying data set and the platform that we’re leveraging has all the right statistics, then this will have solved it. you’re at that point merely retrieving the knowledge that is already embedded in the model, in the model parameters. That’s not, unfortunately, what we see, where there is such detailed and also niche knowledge about people that if you just take one example, it might feel very mundane, but it’s quite rich when you put together, that you do need to do a lot of bespoke data collection to better understand people. And this is also, I think what makes this particular, job fun, which you want to deeply understand people, and the process of deeply understanding them requires a lot of attention to the details. And you do need to pay attention to and pay respect to the daily lives that people lead.

Scaling Simulation: From Thousands to Societies

Vibhu [00:40:04]: I wanna talk about scaling simulation.

Vibhu [00:40:07]: So what can’t we simulate, what can we simulate, and how does scaling affect this? So how big are the models? What if we go from, 8B, like, couple 100 billion

Vibhu [00:40:18]: Like billion000 parameters, billion000? Do we get scaling? Any interesting emergence? Like, at a certain scale, at a certain amount of training, you uncover anything unusual and any learnings from that?

Joon [00:40:31]: What we are seeing is at Simile, so we do post-train our own model. The thing that we’re seeing is the early glimpse of scaling law in simulations. The more data about humans and more compute you ingest, you start to get predictive and predictable gains of the model performance in simulating it, simulating people.

Vibhu [00:40:51]: Ooh. We need a scaling law curve.