Aug 21, 2026 · 1h 11m · latent-space
Simulating Humanity: from Generative Agents to 8 Billion Digital Twins — Joon Sung Park, Simile AI
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth interview, Simile AI co-founder Joon Sung Park discusses the technical evolution, validation, and broad societal applications of generative agents and behavioral foundation models. He explains how simulating authentic human quirks and multi-agent dynamics enables decision-makers to de-risk complex policies and products before real-world execution.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 8.4% of the talking time here. How this is scored →
The hosts as informed peer 4.2 Guest teaching 4.5 Guest disagreement 1.6 The hosts pushing back 1.7
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 40:00 Joon rejects brute-force combinatorial personas
Joon directly challenges the premise that combinatorial persona generation solves simulation, explaining that it merely pulls surface statistics without capturing genuine human nuance.
Hardest push from the hosts ▶ 22:15 Swyx challenges need for simulation on obvious outcomesSwyx presses Joon on why anyone would pay for a simulation if the direction and intuitive effect of a decision are already obvious.
Biggest teaching moment ▶ 11:16 Joon breaks down social physics vs prompt engineeringJoon educates the hosts on why foundational parameter updates are essential when models must learn uncaptured social physics rather than merely reacting to prompts.
The host holds their own ▶ 51:15 Alessio reframes simulation economics against physical studiesAlessio steps in with sharp domain expertise to counter Swyx's cost fears, demonstrating that synthetic trials are orders of magnitude cheaper than physical human studies.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| From Fine Arts to Computer Science Research | 4 | 2 | 1 | 1 | The hosts welcome Joon Sung Park, asking about his artistic background and praising the impactful Generative Agents Smallville paper. Alessio notes the underrated memory component, showing domain awareness while allowing Joon to unpack his research history. | |
| Prioritizing User Simulation Over Personal Assistants | 3 | 4 | 2 | 1 | Swyx asks about alternative ideas explored during their 'time machine' exercise. Joon explains why building accurate models of user behavior must precede automated personal assistants, offering a gentle hot take on current assistant capabilities. | |
| Memory Architecture, Model Weights, and Social Physics | 4 | 5 | 2 | 2 | Alessio asks why changes are needed at the model level rather than just prompt engineering. Joon explains that prompt engineering suffices when base physics exist, but foundational model adjustments are needed to teach models humanity's uncaptured social physics. | |
| Three Data Pillars for Behavioral Foundation Models | 4 | 6 | 1 | 1 | Alessio asks about the specific data pillars needed to build a behavioral foundation model. Joon breaks down qualitative interview data, observational data, and causal RCT data, highlighting that causal mechanisms shape future decisions rather than simple predictions. | |
| Acquiring Behavioral Data with Real Stakes | 3 | 5 | 1 | 2 | Swyx probes how Simile captures nuanced behavioral data if people don't log everything themselves. Joon clarifies their consent-based virtual lab methodology and emphasizes putting real stakes on participant decisions to separate behavioral truth from attitudinal noise. | |
| Commercial Applications and Sensitive Policy Guardrails | 4 | 3 | 2 | 2 | Alessio asks about enterprise customer workflows and customization, while Swyx asks about political applications. Joon describes population modeling use cases and deliberately emphasizes holding off on politics due to safety guardrails. | |
| Simulation vs. Prediction: Uncovering Non-Obvious Trajectories | 5 | 5 | 3 | 3 | Swyx pushes on when a customer truly needs simulation if the general direction of an outcome is already intuitive. Joon refutes the idea that simulation is simple prediction, using Isaac Asimov's Foundation series to illustrate finding non-obvious multi-step trajectories. | |
| Grounding Digital Twins: The 1,000 Agents Validation Study | 5 | 6 | 2 | 2 | Alessio raises the core skepticism of hallucination cascading across agents and asks how digital twins are grounded and evaluated. Joon details their 1,000-person study achieving 85% accuracy and contrasts Simile's imperfect human modeling with frontier models aiming for super-rationality. | |
| Post-Training with Open Science Data and Model Levels | 4 | 5 | 1 | 1 | Joon walks through post-training models using Open Science Foundation pre-registered RCT datasets to overcome publication bias. Alessio asks about individual versus population level training architecture, and Joon explains how Simile handles both. | |
| Modeling Human Quirks and Irrationality in AI | 5 | 5 | 3 | 3 | Swyx brings up Tencent's 1-billion persona paper and asks if prompting combinatorial matrices is sufficient. Joon respectfully rejects the sufficiency of that approach, arguing that synthetic matrix generation only retrieves surface statistics rather than nuanced human behaviors. | |
| Scaling Laws in Simulation and Civilizational Challenges | 4 | 5 | 1 | 1 | Alessio asks whether scaling laws emerge in multi-agent simulations as parameter sizes and data grow. Joon confirms early scaling law indicators and lays out a 10-year vision of simulating 8 billion people to solve wicked coordination problems like climate change. | |
| From Schelling's Grid Models to Generative Agent Simulations | 4 | 5 | 1 | 1 | Joon introduces Thomas Schelling's early agent-based segregation models to demonstrate how generative agents bring rich fidelity to historically simple grid simulations. Swyx adds historical context regarding housing quotas in Singapore. | |
| Economics, Infrastructure, and Cost-Benefit of Simulation Scale | 6 | 4 | 2 | 4 | Swyx expresses concern over the steep computational costs of scaling multi-agent interactions. Alessio counters that synthetic simulation is vastly cheaper than physical real-world experimentation, while Joon frames the investment around high-upside societal decision-making. | |
| Multimodal Digital Twins and Real-World Product Testing | 5 | 5 | 2 | 2 | Alessio asks about sparse architectures and efficiency, while Swyx asks for enterprise case studies. Joon shares details on multimodal product testing with Wealthfront and explains why digital twin panels expand democratic participation in organizational decisions. | |
| Benchmarking the Simulation Industry's Developmental Stage | 4 | 4 | 1 | 1 | Swyx asks about overall market size, and Joon reframes simulation beyond the $100B market research sector to all human decision-making. Joon then connects his figure painting background to simulation as capturing essential human texture. | |
| Testing Long-Term Policy Interventions and Universal Basic Income | 4 | 4 | 2 | 2 | Alessio asks Joon to look 10 years ahead to high-impact simulation targets like Universal Basic Income. Swyx and Joon discuss OpenAI's real-world UBI trial, pointing out that simulations can test varied implementations far cheaper than multi-million dollar field studies. | |
| Simile AI's Team DNA, Co-Founders, and Hiring Vision | 3 | 3 | 0 | 0 | Alessio and Swyx ask about Simile AI's hybrid research-product identity, co-founders (Percy Liang, Michael Bernstein, Laney Yellen), and hiring roadmap. Joon outlines their team culture and recruitment across engineering and research. |
Statements from this episode (27)
Park: LLMs Trained on Web Data Can Accurately Simulate Human Behavior
“These models are actually trained on this very broad data from the web, right? So these are human behavioral data. It's the, it's social media, Wikipedia, all these kind of data. So if you poke At the right angle, then you could see human behavior that would j…”
Park: Accurate human simulation models must precede complex automation agents
“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.”
Park: No true AI personal assistant exists today despite industry ambitions
“My hot take actually here though, is I don't think we've actually seen a true personal assistant that's actually useful in ways that actually meets the ambition of that particular line of work.”
Park: Training is required for new world physics, while prompting suffices for reactions
“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 …”
Park: Public LLMs lack real human social physics due to web data bias
“I don't think the model has yet, at least the models that are out in the open, has yet learned the complete mapping of social physics of humanity. This actually is one of the core thesis of simile, right? And one of the core reason why that is the case is if y…”
Park: Simile uses life story interviews for behavioral models
“We think about data in three buckets. So one bucket is actually, we interview data, for instance. It's quite interesting. A qualitative, rich qualitative data is interesting. It's not behavioral, but we would literally ask people, hey, tell me the story of you…”
Park: Childhood memories and trauma inform behavioral foundation models
“Even understanding their childhood memory or Even their trauma, their first love, these kind of things, quite informative in ways that's really hard to predict.”
Park: Decision-makers want simulation to shape the future, not predict it
“The reason why people are interested in simulation actually isn't because they want to predict the future. If you're winning against, if you're trying to win against a stock market, predicting the future is interesting, but most people, most decision makers, w…”
Park: Real decision stakes distinguish behavioral data from attitudinal data
“What makes the difference between what is attitudinal versus behavioral is if the stake in your decision is real. That's ultimately what makes it behavioral.”
Park: Simile AI has a strategic partnership with Gallup
“We do have a strategic partnership with Gallup, and of course, Gallup is deep into policy space and so forth.”
Park: Advanced simulation discovers multi-step paths to achieve a target 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 t…”
Park: Generative agent digital twins replicate human behavior at 85% accuracy
“And this is where we basically could replicate people's behaviors and attitudes, 85% as accurately as people would replicate their own. So that actually was the first really paper that gave this validated results that we can actually model individuals in an ac…”
Park: Simile aims to build models that replicate human errors
“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 kind of mistake.”
Park: Frontier models hit only 20-30% accuracy predicting niche human behavior
“Where in some cases, the model performance of frontier models go all the way down to 20, 30%. Especially if you go into that more niche population on topics that our customers will actually care about. On more gen pop, it might be around 50 to 60%.”
Park: Training on randomized controlled trials improves AI human-behavior prediction
“That by collecting a lot of these randomized control trials that are really well designed, we can make significant improvement in models capability to predict human behaviors.”
Park: Simile always trains separate population-level and individual-level models
“And this is actually what we end up doing at Simile II. We always train two distinct models. One is what we call the population level model. The other is what we call the individual level model.”
Park: Facebook data better captures human baseline state than LinkedIn or Twitter
“If we were to look at purely social media, like if you really, you know, if I were, you know, if I had to really pick, Facebook likely is interesting because I actually do think it is most sort of a default version of people because you go to LinkedIn, it's ve…”
Park: Simile observes empirical scaling laws when modeling human behavior
“What we are seeing is at Simile, so we do post-train our own model. The thing that we're actually seeing is the early glimpse of scaling law in simulations. The more data about humans and more compute you ingest, you actually start to get predictive and predic…”
Park: A Nobel Prize in economics will be won from societal simulations
“And, you know, I also think, yes, I mean, I think there's a Nobel Prize to be won there, which wouldn't be surprising.”
Park: Simile collects weekly data on tens of thousands and accesses tens of millions
“Today what we do is every week we are collecting data on the scale of tens of thousands people's data. And we actually have panel partnerships that gets us to tens of millions of people globally.”
Park: Future simulations will cost as much to run as training models
“My hunch here is I do think in the next Some number of years, we will start creating simulations that will actually cost as much as training a foundation model.”
Park: Wealthfront was an early customer testing products with Simile
“Worldfront is an interesting one because one of the things they were trying to do, they were one of the first customers that wanted to actually do product testing. That goes beyond just asking people what they think about, let's say, behavior experiments and s…”
Park: Simile AI agents can autonomously navigate live website URLs
“Some of the things that our agents can also do is you can be given a domain, like a website URL and actually go use it for a while.”
Park: Agent simulations are replacing traditional human panel research
“Today, a lot of the demand does come from basically, like, the places where people have historically used human panels, we can basically now replace with agents. And these synthetic populations.”
Park: Agent simulations beat multi-year field trials via instant repeated execution
“This is the reason why you want to run a simulation. You spend five years, forty million dollars on this one study and have one finding. But if you can run simulation many, many times instantly, then that's the value.”