Talks

Simulating Humanity: from Generative Agents to 8 Billion Digital Twins — Joon Sung Park

43 min read

Joon Sung Park · 2026 · 1:11:01 · ▶ Watch

Summary

Central question. Can you build a faithful, queryable copy of human behaviour, first of a few thousand people and eventually of everyone, and use it to test decisions before making them in the real world?

Who is talking. Joon Sung Park is the lead author of the 2023 “Generative Agents” paper (the Smallville experiment, where language-model characters remembered, planned and gossiped in a sandbox town). He is now co-founder and CEO of Simile, a start-up that sells “digital twins” of consumer populations. Two podcast hosts interview him. His 2023 work is the earlier chapter of this story. This conversation is the later one: what happened when the research became a company.

Main thesis.

  1. Frontier chat models are trained to be rational, helpful reasoners. They are poor simulators of real, inconsistent people. Prompting them with a persona is not enough; you must change their weights with data about what people actually do.
  2. The data that matters comes in three kinds: long life-story interviews, observational and transaction records, and, above all, randomized controlled trials, because only experiments show cause and mechanism.
  3. The goal is not to predict the future but to find a path to a chosen future. “Simulation” differs from “prediction” because it shows the steps, and the best step may be counterintuitive.
  4. His headline evidence is a 2024 study of 1,052 people in which interview-based agents matched participants’ survey answers “85% as accurately as people would replicate their own”. Later, he says, post-training on experiment data raises accuracy further and shows early “scaling laws” for simulation.
  5. The long ambition is a society-scale simulation (8 billion agents) to study wicked problems such as climate coordination, democratic collapse and basic income. He compares the work to painting: a good model is not a perfect copy but a picture that surfaces what matters.

Close. He answers the “are we in a simulation?” question with a shrug (“I don’t think that makes our experience any less real”) and ends with a hiring pitch.

Stance. An enthusiastic founder-researcher. He is candid about limits in places (he says frontier models drop to 20 to 30 percent accuracy on niche groups, and that Simile is careful about politics), but the interview is friendly. Hosts do raise real challenges: how do you ground it, and do you need it if the result is obvious? This page keeps the company’s claims apart from the published evidence in Part 1 and in idea 4.

Map

#Point
1Hook: simulate 8 billion people; climate change as a “wicked problem”
2Park’s path: Korea, Boston, painting, then computation as an artist’s medium
3Foundation models as “stem cells”; the time machine game; “let’s just create a world”
4Runner-up idea: personal agents. A good assistant needs a good model of the user. Markdown-file memory
5Why prompting is not enough: “social physics” must be learned; web data is attitudinal, not behavioural
6Three data buckets: interviews, observational data, randomized controlled trials
7Prediction vs shaping the future; getting behavioural data with real stakes
8How customers use it: population filter plus an “environment”, concept testing, focus groups, earnings calls. Politics held back
9”When do I need a simulation if I know the answer?” Magnitude, and the Foundation (psychohistory) example
10Grounding: the 1,000-person study, 85%, and why frontier models fall to 20 to 60%
11Post-training on randomized trials from the Open Science platform
12Population-level vs individual-level models; what is “fundamentally human”; which social data to want
13The one-billion-personas approach, and why bespoke data is still needed
14Early scaling laws; the 8-billion-agent vision; wicked problems; a Nobel to be won
15Schelling’s segregation model and the return of agent-based modelling
16Cost: tens of thousands of people weekly, reuse of twins, a simulation that costs as much as a foundation model
17Customers (CVS, Gallup, Deloitte, Wealthfront), multimodal agents using websites, “not a tool for market research”, market size
18Where the field stands; simulation as painting; the fractal; Studs Terkel’s Working
19UBI as a test case; “you spend 5 years, $40 million on this one study”
20”Are we in a simulation?” Close; company, team and hiring

Putting it together

The mental model in one paragraph. A language model trained on the web has read what people say. A simulator of people has to know what they do, and why, which requires data the web mostly lacks: private life stories, transactions, and experiments in which one variable changes and behaviour moves. Park’s bet is that you can turn that data into a model that is deliberately less rational than a chat assistant and that you can then (a) query it like a survey panel today, (b) run it as many interacting agents to study emergent effects tomorrow, and (c) use it to search for paths to outcomes, not only to forecast them. The strongest published evidence is that interview-grounded agents track individual survey answers and experiment results fairly well on a U.S. sample. The weakest link is the jump from “tracks survey answers on average” to “a trustworthy stand-in for a group you have not measured”, and from there to “a society”.

Three tensions to keep open

  1. Replace or inform? Park says the tool is “obviously not replacing human panel” while also saying it takes over work “where people have historically used human panels”. Both are true only if each use is grounded by fresh human data.
  2. Average accuracy vs the decision at hand. A score of 85% (normalized) is an average over many questions and people. A firm launching one product needs to know whether its answer is among the good ones. Independent tests (idea 4) show errors concentrate in some subgroups and in variance.
  3. Prediction vs causation. Simulation is sold as the tool for “what do I do?” That is a causal claim, and a model that is accurate at describing answers is not automatically accurate at responding to an intervention. Park’s answer (train on randomized trials) is the right kind, but the public evidence is thin.

Reading path

  • Start: Park et al., Generative Agent Simulations of 1,000 People (2024), then the Pew Research Center test, “Can AI Stand In for Human Survey-Takers? Not Really” (2026).
  • Then: Thomas Schelling, Micromotives and Macrobehavior (1978) or the segregation chapter; Park et al., Generative Agents (2023).
  • Deeper: Peng et al., Digital Twins as Funhouse Mirrors (2025); Argyle et al., “Out of One, Many” (2023)†; Epstein and Axtell, Growing Artificial Societies (1996)†.

Part 1: Before you read

Key terms

  • Generative agent. A language-model character with a memory stream, the ability to reflect on it, and a plan. Introduced in the 2023 Smallville paper. Park notes the term “agent” barely existed then.
  • Digital twin (of a person). In this talk, an agent built from one real person’s data (an interview plus survey and behaviour records) so that it answers as that person would. Careful: in engineering, “digital twin” means a live model of a machine. Here it means a model of a human.
  • Synthetic respondent / synthetic panel. An AI-generated stand-in for a survey or panel participant.
  • Population-level vs individual-level model. Simile trains both. One takes a description of a group plus a stimulus and returns a distribution; the other predicts one person.
  • Attitudinal vs behavioural data. What people say they think vs what they do with something at stake. Park’s test: “the stake in your decision is real”.
  • Social physics. His phrase for the regularities of how people behave, which he says public models “[have not] learned the complete mapping of”. Note: the same phrase has older meanings (Comte; Alex Pentland’s 2014 book of that title)†.
  • Normalized accuracy. The paper’s metric: a model’s accuracy divided by how well each person reproduces their own answers two weeks later. 1.0 means “as good as the person is at repeating themselves”. It is not the same as “85 percent of answers right”.
  • RCT (randomized controlled trial). People are randomly assigned to conditions that differ in one variable, so that differences in outcome show cause.
  • Concept testing. Showing messages, products or ideas to a target group before launch, a standard marketing task.
  • Psychohistory. The fictional science in Isaac Asimov’s Foundation, where mathematics predicts the course of a galactic empire and lets a small group steer it.
  • Agent-based model. A simulation made of many simple agents following local rules, studied for the large-scale patterns that appear.

People and works named (a reading list)

Person / workWhy it comes upStart with
Joon Sung Park, Michael Bernstein, Percy Liang, Lainie YallenThe four Simile co-founders he namesPark et al., Generative Agents (UIST 2023); Generative Agent Simulations of 1,000 People (2024)
Percy Liang et al.”Opportunities and Risks of Foundation Models” (2021), where the term “foundation model” was coinedThe report’s introduction
American Voices ProjectSource of the interview scriptStanford Poverty and Inequality Center pages
Thomas SchellingSegregation model; early agent-based modellingMicromotives and Macrobehavior (1978)
Isaac AsimovFoundation, psychohistoryFoundation (1951)
Mikhail Parakhin (Shopify)SimGym, a shopping-journey simulator, raised by a hostHis Latent Space episode
”One billion personas” paperPersona-matrix approach to synthetic peopleTencent’s Persona Hub paper†
Studs Terkel, Working (1974)Raised by a host as a model for deep portraits of ordinary livesWorking

Park’s own frame

Park trained as a painter and came to research through computing as an artist’s medium. His thesis is that behaviour is a modelling problem to be solved with data, not only with bigger language models, and that simulation is closer to portraiture than to forecasting.

The current debate (2025 to 2026)

This section is background, not part of the talk. It separates what the company says from what has been published.

Company claims.

  • Funding and customers. Simile raised a $100 million Series A led by Index Ventures (StreetInsider) and, on 30 July 2026, a $200 million Series B at a $2 billion valuation, led by Greenoaks according to most reports (The Next Web). Named customers include CVS Health, Deloitte and Wealthfront.
  • Accuracy. The company says its answers are “85% to 99% accurate, depending on the group”. The Next Web notes plainly that these figures are Simile’s own. The episode page repeats an “85 to 99%” range against human focus groups but does not explain how it is measured (Latent Space).
  • In the talk, Park adds that frontier models score “20 30%” on niche groups and “50 to 60%” on the general population. These are his statements and are not independently tested in anything cited here.

Published evidence, supporting.

  • The 1,000-person paper (arXiv 2411.10109): 1,052 U.S. participants, stratified on age, region, education, ethnicity, gender, income, neighbourhood and political ideology, each interviewed for about two hours. Interview-based agents reached a normalized accuracy of 0.85 on the General Social Survey (raw accuracy 68.85% against participants’ own two-week repeat accuracy of 81.25%), against 0.71 for agents given only demographics and 0.70 for persona-style agents. On the Big Five personality inventory the normalized correlation was 0.80, and on economic games 0.66. In five replicated social-science experiments, both the participants and the agents replicated the same four and failed on the same fifth, and the effect sizes correlated at r = 0.98. The authors report that the architecture reduces accuracy gaps across racial and ideological groups compared with the demographic baseline. Core analyses were pre-registered. Note that a newer version (v3, June 2026) is listed, so figures may have been updated.

Published evidence, cautionary.

  • Digital Twins as Funhouse Mirrors (Peng et al.), built on the Twin-2K-500 dataset of 2,058 U.S. participants answering more than 500 questions: across 19 pre-registered studies and 164 outcomes, twins were only slightly better than a generic LLM persona at predicting individuals, and showed five systematic distortions, including stereotyping, insufficient individuation and being more rational than the humans (arXiv 2509.19088; dataset: arXiv 2505.17479). These twins were prompted with the data, not post-trained as Simile’s are, so this is a different method from the one Park describes.
  • Pew Research Center (30 September 2026) built AI personas from members of its probability-based American Trends Panel and compared the model’s answers with the real panelists’ on three 2026 surveys. It reports poor estimates of public opinion overall, with the largest errors among Republicans and Black adults (Pew). The model tested is not Simile’s, and the personas used a demographic and typology summary, not a two-hour interview.
  • Studies of synthetic respondents report compressed variance, over-agreement and inflated positive attitudes: see Kaiser et al. in NIM Marketing Intelligence Review (NIM) and a psychometric comparison (arXiv 2601.14264).
  • Methodological objection from industry: standard validation statistics rest on assumptions that do not hold for LLM output, and more synthetic respondents only tighten estimates around a distribution that may already be wrong (Sawtooth Software).

Where that leaves the debate. Evidence supports narrow claims: interview-grounded agents can track survey answers and published experiment effects in one U.S. sample, better than demographics alone. It does not yet support broad claims: that results transfer to other countries, fresh products, or subgroups; that they capture variance and rare views; or that the agents respond correctly to interventions that the training data did not contain. Simile’s higher figures are not yet backed by a public study of the same kind.

The ideas

Park covers a lot in an interview setting: a research origin story, a technical method, a product, a market, a society-scale vision, and a view of art. The traditions behind it meet like this.

 AGENT & COGNITIVE ARCHITECTURES ──┐                ┌── BEHAVIOURAL SCIENCE
 Newell/Simon · Bates · The Sims   │                │   Simon · Kahneman · revealed preference
                                   ▼                ▼
                         ┌─────────────────────────────────┐
 CAUSAL INFERENCE ─────► │  PARK: model what people DO,    │ ◄──── SYNTHETIC RESPONDENTS
 Fisher · Rubin · Pearl  │  then simulate what happens if  │       Argyle · Aher · Wang · Pew
 Lucas critique          └─────────────────────────────────┘
                          ▲              ▲               ▲
        MARKET RESEARCH   │   AGENT-BASED MODELLING      │  MODELS AS REPRESENTATION
        focus groups ·    │   Schelling · Epstein/Axtell │  Box · Borges · Bostrom
        panels · TAM      │   Asimov (psychohistory)     │
  • Where foundation models and “believable agents” come from: idea 1
  • Why chat models are poor people-simulators, and the say-do gap: idea 2
  • Data buckets, experiments and prediction vs shaping the future: idea 3
  • The 1,000-person study and the synthetic-respondent debate: idea 4
  • Product, customers and market research: idea 5
  • Scaling, Schelling, and a society of agents: idea 6
  • Simulation as painting; the simulation hypothesis: idea 7

1. From foundation models to Smallville: “let’s just create a world”

In short. Park says foundation models arrived as a general-purpose “stem cell”. Because they are trained on human-written data, human behaviour “pops out” when you prompt them well. His group played a “time machine game”, asking what single application would matter most ten years on, and answered: recreate the world. A personal assistant came second, and he argues it needs a model of the user first.

🎙 In the talk.

  • On the models: “It was like a stem cell if you were to take a biology analogy”. They are “trained on this very broad data from the web” and “these are human behavioral data”.
  • On the insight: “if you poke at the right angle, then you could see human behavior that would just pop out”.
  • On the ambition: “let’s just create a world”. It was a close second to personal assistants that “do things for you”.
  • On why simulation first: “That’s the basis of our social connection”, by which he means that a friend’s mental model of you is what makes help useful. His example is a food-ordering assistant that must know you dislike pineapple on pizza.
  • On memory design: the team considered a knowledge graph or a bespoke model, then chose “just put everything in a markdown file or text file you’re done”. He sees the same intuition in current personal-agent tools, and also a limit: retrieval over very large memory “takes a lot of work”.
  • On the limit of prompting: “there are certain things you just cannot shape just by prompting the model”.

📚 Deeper theory

  • What Smallville showed. In the 2023 paper, 25 agents in a sandbox town used a memory stream, periodic reflection and planning, and produced emergent social events such as a spontaneous party invitation spreading. The three-part architecture (memory, reflection, planning) is the part Park’s host called “underrated”. The sibling idea in this talk is that the same memory-in-text trick is now standard in agent tools.
  • Why “behaviour pops out” is a strong claim and a weak one. Strong: training text contains vast records of people talking. Weak: text is a performance of behaviour (see idea 2). It contains what people write, not a record of what they do unobserved.
  • Personal assistant vs simulator. The argument that a simulator should come first is a user-modelling argument: an agent that acts for you needs a model that predicts your preferences, including unstated ones. This is what recommender systems do in a thinner form.

📚 Roots: Believable agents and cognitive architectures

The problem. How do you make a computer character that behaves in ways people find plausible over long periods, without scripting every case?

The point. Smallville is the first time that a general language model, not a hand-built cognitive model, was enough to make that plausible. Park’s agents replace the rule-heavy architectures of earlier decades with text, memory and retrieval.

The chain of thinkers

  1. Allen Newell and Herbert Simon, from the 1950s to 1970s. General Problem Solver; the idea that intelligence is symbol manipulation with goals and memory. Their “cognitive architecture” tradition (later Soar, and ACT-R from John Anderson) aims to model how human minds work by explicit mechanism†.
  2. Joseph Weizenbaum, ELIZA (1966). A scripted chatbot that people treated as a person: early evidence of how readily we attribute a mind to text.
  3. Joseph Bates, “The Role of Emotion in Believable Agents” (1994). The term “believable agents” in interactive art and games: a character does not have to be intelligent to be convincing, but it must have coherent personality and emotion†.
  4. Will Wright, The Sims (2000). Mass-market simulated people with needs and routines. Smallville is often described as a Sims-like town driven by language models rather than hand-written needs†.
  5. Foundation models (2021) and in-context agents (2022 to 2023). The Opportunities and Risks report, then chain-of-thought and tool-use work, gave the models the means to plan and reflect.

Where others push back. Believability is not fidelity. A character can be convincing and wrong about real people. This is the gap the rest of the talk addresses with data.

What they leave out. The 2023 agents were fictional characters. Moving to real, specific people (idea 4) changes the question from “is this plausible?” to “is this correct?“.

Check yourself

  • Why does Park think a personal assistant needs a user model before it needs tools?
  • What did the team decide to use for memory, and what limit does he name?
  • What is the difference between a believable character and an accurate simulation of a real person?

2. Social physics: why chat models are poor people-simulators

In short. Park argues that models trained for reasoning have learned what people say online and what experts consider right, not the full pattern of how people actually behave. To simulate people, a model must learn this “social physics”, and that needs training on new data about behaviour, not prompts. He adds that a good simulator must be as fallible as its subjects.

🎙 In the talk.

  • On when training is needed: he says a model can simply be prompted when it already has “the physics” and only has to react to an environment; but the open models have not “learned the complete mapping of social physics of humanity”.
  • On why: web data is “the selfexposed attitudinal data with some behavioral data that sprinkle around here and there”. What is missing is “the dark knowledge of humanity that we haven’t quite captured”.
  • On the target: current frontier models are tuned to become “the super rational objective machines” through expert data; “the models that we’re talking about here … are models that are as dumb as I am”.
  • On the example: he used to walk 40 minutes home from campus because it helped him think. A model would suggest a bus. “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”.
  • A host links this to Moravec’s paradox (caption: “VX paradox”), and Park agrees: simulating human errors is harder than simulating expert reasoning.

📚 Deeper theory

  • The say-do gap. Economists distinguish stated from revealed preference: what people say they would pay vs what they pay. Surveys measure the first. A large literature in marketing and public health finds big gaps between stated intentions and later behaviour†. Park’s “attitudinal vs behavioural” split is this old point applied to training data.
  • Why rational models make bad human models. If a model is trained to be correct and helpful, it will tend to give the sensible answer. Real people procrastinate, are loss-averse, ignore base rates and are swayed by framing. The Funhouse Mirrors study finds the same: twins “more rational than their human counterparts” (arXiv).
  • Moravec’s paradox. Hans Moravec (1988) noted that what is hard for humans (logic, chess) is easy for computers, and what is easy for humans (perception, motor skill) is hard†. The host applies it by analogy: being irrational in the right ways is “easy” for people and hard for models.
  • Prompting vs weights. Park’s practical rule: if the model already contains the distribution you need, prompting can retrieve it. If not, retrieval cannot invent it. Whether the base model lacks the distribution is an empirical question, and the 20 to 30 percent figure is his answer for niche groups.

📚 Roots: Behavioural science and the say-do gap

The problem. How do people actually decide, as opposed to how a rational agent would?

The point. If economic and psychological theory says real choices depart systematically from rationality, then a simulator tuned for rationality will be systematically wrong in the same direction. Park’s “as dumb as I am” is a modelling requirement.

The chain of thinkers

  1. Herbert Simon, bounded rationality (1950s). People satisfice under limits of information and attention.
  2. Paul Samuelson, revealed preference (1938). Preferences are inferred from choices, not statements†.
  3. Daniel Kahneman and Amos Tversky, heuristics and biases; prospect theory (1970s to 1979). Systematic, predictable departures from rational choice: framing, loss aversion, anchoring.
  4. Richard Thaler and behavioural economics (1980s on). Predictable irrationality becomes a basis for policy (“nudges”).
  5. Survey methodology on social desirability. Respondents give answers that look good. A model trained on public text inherits this.

Where others push back. Skeptics ask whether rational-looking text really is the problem, or whether the problem is variance: a model that answers like the average person will under-represent how much people differ. Studies of synthetic respondents report exactly that compression of variance (Part 1).

What they leave out. Park does not say how Simile checks that its post-trained model has become appropriately irrational for a given group, beyond the aggregate benchmarks in idea 4.

Check yourself

  • What does Park mean by “attitudinal” data and why is web text mostly that?
  • Why might a model trained to be helpful be a bad simulator?
  • What is the difference between a model lacking “social physics” and a model that has it but is not prompted correctly?

3. Three kinds of data, and prediction vs shaping the future

In short. Simile trains on interviews, observational data and randomized trials. Park calls the last “perhaps the most important” because it shows cause and mechanism. He says customers do not mostly want a forecast; they want to know what to do. That is why simulation, as he defines it, differs from prediction.

🎙 In the talk.

  • The buckets: rich interviews (“tell me the story of your life”) that give a lot of texture; observational data such as transactions or scraped web behaviour, which “give you the base statistics”; and trials that describe “the cause and mechanism”.
  • On the coffee example: compare the day you drank coffee with the day you did not, to see whether behaviour changes.
  • On shaping: “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 is going to tank in two quarters”.
  • On why trial data is rare: “the world is our ground truth, but it happens once”.
  • On stakes: “what makes the difference between what is attitudinal versus behavioral is if the stake in your decision is real”. Participants in a store experiment actually get the items they pick.
  • On consent: people are invited to “be a member of this community” to share data and have themselves represented.
  • On simulation vs prediction: “this is the reason why simulation is actually different from prediction”. In its highest form you give it “not a problem or question like what would people answer to the survey” but a goal, and it finds the path.
  • The Foundation example: the scientists find the empire will fall; the simulator’s first move is to exile them to Terminus, “so counterintuitive”. His market example: a promotion that raises EV sales might lower perception of the firm’s other cars.
  • A host connects this to Shopify’s SimGym, “interventions on a shopping trajectory”, and says it is “not about the near-term direction” but “multiple turns of interactions”.

📚 Deeper theory

  • Why trials are special. An observational record shows what happened. A randomized trial shows what happens because of a change, since random assignment breaks the link between who gets the change and their other traits. Causal questions need this or strong assumptions.
  • The Lucas critique. Economists warn that a model fitted to past behaviour may fail when a policy changes the rules, because people change how they behave (Robert Lucas, 1976)†. A simulator that learns from past data must show it responds correctly to new interventions. Training on randomized trials is a direct answer, but only for the kinds of interventions the trials cover.
  • Path-finding as optimisation. “Find the path to a goal” means searching over interventions inside a model. The result is only as reliable as the model’s causal fidelity, and optimisation tends to find the model’s errors first. The Foundation story shows a fictional model that is right; real models need validation on out-of-sample interventions.
  • Consent and recruitment. The talk describes opt-in, incentivised panels. Opt-in samples can differ from the population in ways that matter (a point made in the survey literature). The 1,000-person study mitigated this with a stratified recruit.

📚 Roots: Causal inference and experimentation

The problem. How can you learn the effect of doing X when you can only ever see one world?

The point. Park’s third data bucket is the experimental tradition. His claim is that training on many designed experiments teaches a model mechanism and not only correlation.

The chain of thinkers

  1. Ronald Fisher, randomization and design of experiments (1920s to 1935). Random assignment as the basis of valid inference.
  2. Donald Rubin, potential outcomes (1974). Each person has an outcome under each condition; we observe only one. This is Park’s “it happens once”.
  3. Judea Pearl, causal diagrams (1990s to 2000). Distinguishes seeing from doing and from imagining (“counterfactuals”), exactly the three things a policy simulator wants.
  4. Robert Lucas, the Lucas critique (1976). Policy changes alter behaviour, so models of past behaviour mislead†.
  5. The replication crisis and pre-registration (2010s). Park describes the Open Science platform where scientists register hypotheses before running studies. Pre-registration is a defence against publishing only the lucky results.

Where others push back. (a) Published experiments are mostly small, on students or online samples, in a few countries. (b) A model that fits many experiments may still fail on a new one. (c) Counterfactual claims about society-level change cannot be tested with a single experiment.

What they leave out. How well post-trained models predict held-out experiments, and in which populations. The talk reports no numbers for that follow-up study.

Check yourself

  • Why does Park say randomized trials are the most important data bucket?
  • What is the difference between predicting sales and finding a path to avoid a sales fall?
  • What would a Lucas-critique test of a policy simulator look like?

4. Evidence: the 1,000-person study and the claim of 85%

In short. To show that digital twins are grounded, Park’s group recruited 1,000 representative Americans, interviewed them, built twins, and tested the twins against the same people two weeks later. The headline is 85%. Park then argues frontier models alone fall far below that on the groups customers care about, and that post-training on trials improves things further. This idea sets out what the study did and did not show.

🎙 In the talk.

  • The host’s challenge: “LLM hallucinate, you’re just hallucinating layer after layer”.
  • Park’s design: a representative U.S. sample, two hours of data collection “paired with a lot of behavioral data”, interview script from the American Voices Project, then people go away “for a couple of weeks” and return for surveys, experiments and economic games. Twins “predict how the source individuals would have acted”.
  • The result: “could replicate people’s behaviors and attitudes 85% as accurately as people would replicate their own”.
  • On gaps: frontier models “go all the way down to 20 30%” for niche populations and “around 50 to 60%” for the general population; “So it’s not very robust”.
  • On the follow-up: post-training on randomized trials from a public pre-registration platform “made a point” that a lot of well-designed trials can “make significant improvement”. This model “is not something that we’re serving commercially”.
  • On the tuning: both population-level and individual-level models are trained from similar inputs.
  • On the one-billion-personas approach: it relies on “known statistics that went into training the model”; if it worked “we actually have solve simulation”, but “that’s not unfortunately what we see” and “you actually do need to do a lot of bespoke data collection”.

📚 Deeper theory

  • What 85% means. In the paper, the number is a ratio: agents’ raw accuracy on General Social Survey items (68.85%) divided by participants’ own accuracy when re-answering after two weeks (81.25%) gives 0.85. The ratio corrects for the fact that people are not perfectly consistent, which is sensible. It also means the ceiling is not 100%, and “85% as accurately” does not mean “wrong 15% of the time”. Figures from the paper’s text; see Part 1.
  • What it did not show. One U.S. sample, English interviews, a two-week gap, mostly survey-style outcomes, and the agents saw rich interview data per person. A client’s twins are built from different, usually thinner, inputs and for new topics. The paper’s own baselines show the interview mattered: demographics alone reached 0.71.
  • Individual vs group. Matching a person is a harder test than matching a group average. But a business often needs the group distribution, including variance, and compressed variance is a recurring failure of synthetic respondents (Part 1).
  • Other tests. The Funhouse Mirrors mega-study, Pew’s test and several variance studies (Part 1) find weaker results with other methods and data. They do not test Simile’s system, so they show that the claim is method-specific, not that it is false.
  • Company numbers vs paper numbers. The 85% is a published, pre-registered result. The “85 to 99%” range used in marketing is not documented publicly in the sources found for this page. Treat them as different kinds of claim.

📚 Roots: Synthetic respondents and “silicon samples”

The problem. Can a language model stand in for a human survey respondent, and how would you know?

The point. Park’s 2024 paper is one of the most-cited answers that “yes, at least for some tasks, if you ground the model in rich individual data”. The critics’ answer is “not reliably, and not for the quantities you care about”.

The chain of thinkers

  1. Lisa Argyle et al., “Out of One, Many” (2023). Showed that GPT-3 conditioned on demographic backstories could reproduce U.S. voting-group response patterns, introducing “silicon samples”†.
  2. Gati Aher, Rosa Arriaga and Adam Kalai, “Using Large Language Models to Simulate Multiple Humans” (2023). “Turing experiments”: replaying classic studies with models†.
  3. Danica Dillion et al., “Can AI language models replace human participants?” (2023). Reported high agreement on moral judgments, with caveats about whose judgments†.
  4. Angelina Wang et al. (2024). Argued that models cannot portray identity groups well, flattening them and speaking about rather than as them. Cited in the 1,000-person paper (arXiv).
  5. Park et al. (2024). Interview-grounded agents, individual-level test, pre-registered.
  6. Peng et al. (2025), Pew Research Center (2026), Kaiser et al. (2026). Stress tests with mixed or negative results (Part 1).

Where others push back. (1) Variance compression: twins give central answers. (2) Subgroup error: the groups that matter for equity are those where models do worst. (3) Validation: matching known survey items is not the same as predicting unknown ones. (4) Sample size does not fix bias in the model.

What they leave out. Park does not discuss variance, subgroup error or non-U.S. performance in this interview. Given Simile’s global panel partnerships, those are the natural next questions, and the answer is not in the public record cited here.

Check yourself

  • Why is the normalized-accuracy ceiling below 100%?
  • Which comparison in the paper shows the interview mattered?
  • Name two reasons a strong result on 1,052 Americans might not transfer to a client’s product launch.

5. The product and the market: concept testing, panels, and “voices in the room”

In short. Today customers choose a population, get twins of it, and ask them questions or run tests. Park says this fits places that used human panels, and that simulation is a decision tool, not market research. He argues the real prize is that more people are “consulted” because asking real people is slow, costly and often impossible.

🎙 In the talk.

  • The workflow: recruit the target group “with consent and with incentives”, collect data, build the model, then query it with “a filter that is a description of the population that you want to talk to” plus “an environment”: a survey, a behavioural experiment or an A/B test.
  • Use cases: “the core use cases are things like concept testing”; also focus groups and “even modeling things like the earnings call for public companies”.
  • On politics: “I do see politics as an area where a company has to be particularly thoughtful about the way they operate and make impact”. They are building guardrails first. A West Wing plot, in which a president tests disclosure of an illness through a poll, is his example of a counterfactual he would simulate “if I could trust it”.
  • The host’s challenge: if I know it will be bad, why simulate? Park: “understanding the magnitude and understanding how acute something is is actually quite difficult”.
  • Customers: Wealthfront wanted “product testing that goes beyond just asking people”, with agents that traverse “Figma mockups or websites” or “go use” a URL. Other names mentioned: CVS, Gallup (a “strategic partnership”), Deloitte.
  • On replacement: demand comes from “the places where people have historically used human panels”; “this is obviously not replacing human panel”.
  • On the purpose: “ensure that the voices of people is always represented in rooms where the decisions for them is made”. Most people would “never” answer a thousand survey questions.
  • On market size: “market research is a hundred billion dollar industry”, but “simulation is not a tool for market research”. The market is every decision “made about human for human”.

📚 Deeper theory

  • Why panels are costly. Recruiting, screening and incentivising representative respondents is slow and expensive, and response rates are falling. Synthetic panels target this pain point.
  • “Not replacing” vs “replacing”. The company line is that twins are grounded in human data and updated with new human data. The risk named by critics is that over time clients substitute synthetic answers for fresh human ones. The two are reconcilable only if a human-calibration step stays in the loop.
  • Customer-reported evidence is limited. The Next Web reports that CVS used 400,000 “agentic twins” to study how patients take medicine (article). No published outcome study of whether decisions made this way were better was found for this page.
  • The “voices in the room” argument. It is a democratic framing of a commercial tool: represent more stakeholders at low cost. The counter is that a simulated voice is a model’s estimate of a voice, and a confident wrong estimate can crowd out the real person.

📚 Roots: Market research and its measurement problems

The problem. How does an organisation learn what a population wants before it commits money?

The point. Simile enters a long line of tools that trade cost against fidelity: panels, focus groups, conjoint analysis, and now twins.

The chain of thinkers and practices

  1. Robert Merton and the focused interview (1940s) and the later commercial focus group†.
  2. Survey sampling and probability panels (mid-20th century). The standard of representativeness that Pew’s American Trends Panel follows.
  3. Stated-preference methods and conjoint analysis (1970s on). Ask people to trade off features; infer value.
  4. Online opt-in panels (2000s). Cheap, fast, but with data-quality and representativeness concerns (the 1,000-person paper cites such work).
  5. Synthetic panels (2023 on). Model-generated respondents, with vendors, academics and incumbents all publishing claims and critiques.

Where others push back. Industry statisticians argue that the validation methods used for human samples do not apply (Sawtooth link in Part 1). Another line: a market is about new things people have not seen, and a model trained on the past may be weakest exactly there.

What they leave out. Pricing and its link to panel size are mentioned by the press but not in the talk. Park says only that value is argued by comparing the cost of bad decisions.

Check yourself

  • What two inputs does a query to the product take?
  • Why does Park say simulation is not market research?
  • What would you need to see to believe that twins improved decisions, and not only cut costs?

6. Scaling, Schelling and a society of 8 billion

In short. Park says Simile sees the “early glimpse of scaling law” for simulation: more human data and compute give predictable gains. The long goal is many-agent societies, which he links to Thomas Schelling’s segregation model. He wants to ask wicked questions (climate, democracy, money) and expects simulations to be expensive but worth it.

🎙 In the talk.

  • Scaling: “the early glimpse of scaling law in simulations”, with “predictive and predictable gains of the model performance” as data and compute increase.
  • The vision: “can we create a simulation of 8 billion people living on earth”. Questions “are fundamentally about emergence”.
  • Examples: climate change as a “wicked problems” case with many actors and competing incentives; signals of collapsing democracy; the origin of the monetary system. “I think there’s a Nobel Prize to be won there”.
  • Schelling: red and blue dots on a grid; an agent moves if too many neighbours differ; “that preference can be very minute, but the very small difference actually causes the society to segregate completely over time”. It “ended up informing housing policies”, mixed-income housing for instance.
  • Why it faded and returns: “red dots and blue dots is not really a rich description of people”, but generative agents allow agent-based models “high fidelity enough to help us make really complex decisions”.
  • Scale today: “every week we are collecting data on the scale of tens of thousands” of people, and panel partnerships reach “tens of millions of people globally”.
  • Reuse and stability: the agents are “domain agnostic”, and traits such as risk tolerance “doesn’t really change over time”.
  • Sample size: people want large samples to “filter down to any population of their interest”, not for statistical guarantees.
  • Cost: “in the next some number of years we will start creating simulations that will actually cost as much as training a foundation model”; if it could solve climate change, “I would raise the money right now just to run that”.
  • Cost relative to the decision: real studies are expensive, and a decision worth $100 million justifies a small simulation spend. Simile thinks about efficiency but does not “overoptimize too early”.
  • Stage: he compares the field to “where Gypty 3.5, Gypty 4 was” (caption error for GPT): powerful enough to do real damage in some verticals, with more progress ahead.

📚 Deeper theory

  • What Schelling showed, and did not. The model shows that mild preferences can sufficiently produce segregation, not that they did. It is a possibility proof about mechanisms. Later empirical work finds that housing discrimination, income and policy also matter†. Use of the model to inform housing policy is how Park describes it; its direct policy effect is contested in the literature†.
  • Emergence vs calibration. Agent-based models are valued for surprising emergent results, but surprise is not evidence. A society simulation needs validation against real macro outcomes, and with one world and no replays that is very hard.
  • Scaling laws for simulation. Scaling laws in language models are measured on well-defined loss. For simulation, the “loss” is agreement with human behaviour on benchmarks; Park reports an early trend without data. Treat as a claim to test.
  • Interaction is the hard part. Predicting a person’s survey answer is one task. Predicting how thousands of agents influence each other over time, with errors compounding, is another. The talk’s strongest evidence is for the first.
  • Cost framing. A single large simulation costing as much as training a frontier model puts it in a different budget class; the question becomes who could judge whether its output is right.

📚 Roots: Agent-based modelling and the science of emergence

The problem. How do large-scale patterns arise from many individual decisions that no one intended?

The point. Park places himself in the agent-based tradition and says generative agents fix its weakness: agents too thin to be realistic.

The chain of thinkers

  1. Thomas Schelling, “Dynamic Models of Segregation” (1971) and Micromotives and Macrobehavior (1978). The checkerboard model. Schelling shared the 2005 Nobel Prize in Economics for work on conflict and cooperation through game theory†, so Park’s wording that the prize recognised simulation is loose (see Source notes).
  2. John von Neumann and cellular automata (1940s to 1950s). Simple local rules produce complex global behaviour†.
  3. Joshua Epstein and Robert Axtell, Growing Artificial Societies (1996). The “Sugarscape” model, an attempt to “grow” social patterns from agent rules†.
  4. The complexity-science tradition (Santa Fe Institute, 1980s on). Emergence, adaptation, networks.
  5. Isaac Asimov, Foundation (1951). Fiction that gave the field its dream: a mathematics of society that supports intervention. Real social science has repeatedly found this harder than the fiction suggests.

Where others push back. (1) Agent-based models are criticised as under-determined: many rule sets fit the same data. (2) Richer agents add parameters and opacity, which makes validation harder, not easier. (3) Language-model agents are conditioned on the same text, so they may share the same biases and agree too much, which would reduce the diversity that drives emergence.

What they leave out. No example is given of a society-scale result validated against reality. The talk is explicit that the 8-billion simulation is a vision (“the vision”), not a current product.

Check yourself

  • What does Schelling’s model show, and what does it not show?
  • Why might language-model agents be worse at producing diversity than red and blue dots?
  • Why is “one world, no replays” a problem for validating a society simulation?

7. Simulation as painting, and “are we in a simulation?”

In short. Park trained as a realist painter and says simulation is like painting: no portrait is perfect, but a good one shows what matters about its subject. He sees simulation as a way to learn about human society, as AGI research has led people to think about human intelligence. He closes by treating the simulation hypothesis as no reason to doubt experience.

🎙 In the talk.

  • “Simulation is a lot like painting, right?” and the best paintings “teach you something deep about the subject”; none is perfect, but each tries “to highlight the thing that matters the most about the subject”.
  • His subject as a painter was “the more mundane aspect of human lives”. He uses a fractal image: “you either go outward … or you go inward”, and the mundane, taken together, “teaches you something really deep”.
  • “Simulation is really an exercise of understanding more about human society and our collective lives”.
  • A host links this to biographers who follow a subject for a week and to Studs Terkel’s Working.
  • UBI as a test case: “do we need UBI?“. A host says a funded study was done; Park: “You spend 5 years, $40 million on this one study and have one finding. But if you can run simulation many many times instantly, then that’s the value”. He allows that people “want to verify what you think”.
  • The simulation hypothesis: “whether we are in a simulation or not that I don’t think that makes our experience any less real”. The other host offers the numbers argument, then notes it is “computationally very expensive” to be a civilization.
  • His last note: in advanced civilizations of science fiction, there are two pillars, AGI and simulation.

📚 Deeper theory

  • Models as representations. The statistician George Box’s line “all models are wrong, but some are useful” is the philosopher’s version of Park’s painting analogy†. Borges’s one-paragraph story of a map as large as the empire makes the opposite point: a perfect copy is useless†.
  • What the analogy hides. A portrait is judged by a viewer who knows the subject. A simulation used for decisions is judged by outcomes, which come later and which a company may not publish. A painter does not rank subjects by an accuracy score.
  • UBI as a hard test. The study the hosts refer to is probably the OpenResearch cash-transfer study, which gave $1,000 a month to about 3,000 people in Texas and Illinois for three years (NBER working paper). Its authors report a modest reduction in work and not a simple “no”. A simulator can only rehearse a policy as well as its model of labour supply, family structure and local labour markets, and would need to have predicted this in advance to count as a test.
  • The simulation hypothesis. Nick Bostrom’s 2003 argument is a trilemma about posthuman civilizations running ancestor simulations†. The host’s “numbers” argument is its popular form.

📚 Roots: Models, maps and simulation

The problem. What does it mean for a model to be “faithful”, and what is it for?

The point. Park’s painting analogy is a theory of fidelity: selective, interpretive and judged by insight. That differs from the accuracy-benchmark view that dominates the rest of the interview, and the interview moves between the two without comment.

The chain of thinkers

  1. Jorge Luis Borges, “On Exactitude in Science” (1946). A map the size of the territory.
  2. Alfred Korzybski, “the map is not the territory” (1931)†.
  3. George Box (1976, 1987). Models are wrong but useful†.
  4. Jean Baudrillard, Simulacra and Simulation (1981). Copies that replace the original. A thinker Park does not name, but the 2022 precursor paper was called “Social Simulacra” (, caption “social similac”) and the theme runs through the work.
  5. Nick Bostrom, “Are You Living in a Computer Simulation?” (2003).

Where others push back. Insight and accuracy can conflict: a vivid portrait can mislead. For decisions affecting real people, “it taught us something” is a weaker standard than “it was right”.

Check yourself

  • What does the painting analogy imply about how to judge a simulation?
  • Where does Park’s own approach (benchmarks against human answers) pull against it?
  • Why is a UBI study a stronger test of a simulator than a concept test?

Q&A

(Append only if worth keeping.)