# Simile (company)

> Source: https://aiwiki.ai/wiki/simile_ai
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Agents, AI Companies, AI Research, Enterprise AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Simile (company)." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/simile_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Simile** (legal name Simile AI, Inc.) is an American [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company that builds computer simulations of human behavior. It models real people as [AI agents](https://aiwiki.ai/wiki/ai_agents), grounded in interviews, surveys, experiments and behavioral data, and lets businesses ask how customers, employees or whole populations would respond to a product, price, message or policy before it reaches the real world [1][2][23]. The company calls itself "The Simulation Company" and says its mission is "to simulate all eight billion people on earth, accurately and honestly" [2].

Simile was founded by three [Stanford University](https://aiwiki.ai/wiki/stanford_university) researchers, chief executive Joon Sung Park, Michael Bernstein and [Percy Liang](https://aiwiki.ai/wiki/percy_liang), together with Lainie Yallen, whom Park describes as his business counterpart [1][4][12]. Park, Bernstein and Liang were authors of the 2023 "generative agents" paper, which populated a virtual town called Smallville with 25 agents built on a [large language model](https://aiwiki.ai/wiki/large_language_model), and of a 2024 follow-up that built agents from two-hour interviews with 1,052 Americans [14][16]. The company came out of stealth on 12 February 2026 with a $100 million Series A led by [Index Ventures](https://aiwiki.ai/wiki/index_ventures) [3][4], and on 30 July 2026 it announced more than $200 million in new funding at a $2 billion post-money valuation [2][8]. Customers it has named include CVS Health, Wealthfront, Deloitte, Gallup and the Brazilian bank Itaú Unibanco [1][2][26].

| Field | Detail |
|---|---|
| Legal name | Simile AI, Inc. [1][5] |
| Founded | 2025, according to Dealroom [33]; public launch on 12 February 2026 [3] |
| Founders | Joon Sung Park, Michael Bernstein, Percy Liang, Lainie Yallen [4][12] |
| Headquarters | San Francisco, with a smaller office in New York (August 2026); the company started in Palo Alto [2][12] |
| Employees | More than 50 (July 2026) [2]; about 60 (August 2026, per Park) [12] |
| Product | Behavioral simulation platform built on what the company calls "a foundation model for human behavior" [1] |
| Funding | $100 million Series A (February 2026); more than $200 million Series B at a $2 billion post-money valuation (July 2026) [3][2] |
| Website | simile.com (simile.ai redirects there) [1] |

## History

### Research origins at Stanford

Park began a PhD in computer science at Stanford in 2020, advised by Bernstein, who works on social and interactive computing, and Liang, a machine learning and natural language processing researcher [12][13][20][21]. In his first year he was one of the co-authors of "On the Opportunities and Risks of Foundation Models", the 2021 report from Stanford's Center for Research on Foundation Models (CRFM), the center Liang directs, which introduced the term "[foundation models](https://aiwiki.ai/wiki/foundation_models)" [18][21]. Bernstein and Liang were also among its authors [18]. Park has said that the three of them then played what they called a "time machine game", asking which application of the new models would look most important ten years later, and settled on recreating the world in simulation [12].

The first result was "Social Simulacra" (UIST 2022), a prototyping technique in which a language model generates thousands of distinct members of an online community, together with their posts, replies and anti-social behavior, so that designers can see how a social platform might behave once it is populated [15]. In the authors' evaluations, participants were often unable to tell the simulated behavior from real community behavior [15]. Park has said the system was built on [GPT-3](https://aiwiki.ai/wiki/gpt-3), before instruction tuning, and was "very janky" [13].

"Generative Agents: Interactive Simulacra of Human Behavior" followed in April 2023 and appeared at UIST 2023, where it received a best paper award [14][19]. Its [agent memory](https://aiwiki.ai/wiki/agent_memory) architecture extends a language model with a complete natural-language record of an agent's experiences, which the agent periodically synthesizes into higher-level reflections and retrieves to plan what it does next. The authors placed 25 such agents in Smallville, a sandbox town inspired by The Sims. Starting from a single user-specified idea that one agent wanted to throw a Valentine's Day party, the agents spread invitations over the next two days, asked each other on dates to the party and showed up together at the right time [14]. The code was published on GitHub under the Apache 2.0 license in July 2023 and had more than 22,000 stars by October 2026 [29].

According to Park, the conviction to start a company came about half a year after the generative agents paper, when requests arrived first from social scientists who wanted to run experiments on the platform and then from Fortune 500 companies [13]. The team first set out to show that simulations of specific individuals could be accurate [13]. The resulting study, "Generative Agent Simulations of 1,000 People" (November 2024), drew on two-hour interviews with 1,052 Americans, conducted with the American Voices Project interview schedule, and built an agent for each person. The agents reproduced participants' answers on the General Social Survey 85 percent as accurately as the participants reproduced their own answers two weeks later, and they reduced accuracy gaps across racial and ideological groups compared with agents given only demographic descriptions [16]. A revised version posted in 2026 under the title "LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals" reports 83 percent for interview-only agents, 82 percent for survey-only agents, 86 percent for agents built from both, and 74 percent for demographics-only agents, each measured against the participants' two-week test-retest consistency [16].

Park's doctoral dissertation, "Generative Agent Simulations of Human Behavior" (June 2025), received Stanford's Arthur Samuel Award for the best PhD dissertation in computer science [19][1]. Related work by Akaash Kolluri, Shengguang Wu, Park and Bernstein, presented at [EMNLP](https://aiwiki.ai/wiki/emnlp) 2025, [fine-tuned](https://aiwiki.ai/wiki/fine_tuning) language models on SocSci210, a dataset of 2.9 million responses from 400,491 participants in 210 social science experiments [17]. Its strongest model, Socrates-Qwen-14B, produced predictions 26 percent better aligned with human response distributions in unseen studies than its base model, [Qwen2.5](https://aiwiki.ai/wiki/qwen2_5)-14B, and 13 percent better than [GPT-4o](https://aiwiki.ai/wiki/gpt_4o), according to the paper's revised arXiv version; the abstract in the ACL Anthology gives 36 and 15 percent [17].

### Founding and Series A (2025 to February 2026)

Simile spun out of Stanford [6]. Bloomberg reported that it spent about seven months developing a model trained on interviews with real people about their lives, historic transaction data and text from scientific journals on behavioral experiments before it took on a public profile in February 2026 [6]. The startup database Dealroom gives 2025 as the founding year [33]. In February 2026 Simile told Bloomberg that CVS Health had been using the model for five months [6]. Park later said the relationship began after a CVS executive who leads the company's human insights work read the paper validating the agent simulations and was introduced to Park through a cousin [13].

On 12 February 2026 Simile announced "$100M in funding led by Index Ventures, with participation from Hanabi, A*, Bain Capital Ventures, Andrej Karpathy, Fei-Fei Li, Adam D'Angelo, Guillermo Rauch, Scott Belsky, and others" [3]. Latham & Watkins, which advised the company, described the round as a closed $100 million Series A [5]. Bloomberg reported that Simile did not disclose a valuation [6]. In Index's announcement, partner Shardul Shah called Park "an oil painter turned entrepreneur" and wrote that he had not experienced comparable market pull "since the early days of Wiz" [4].

At launch Simile said customers were using it to "rehearse earnings calls, model litigation outcomes, and test policy changes" [3]. Park told Bloomberg Television: "In a recent earnings call, we simulated, we actually predicted, eight out of 10 questions that were actually asked on this call" [6]. SiliconANGLE listed CVS Health and Telstra Group among the first customers [7]. In March 2026 Gizmodo, summarizing a Wall Street Journal profile, reported that a CVS executive had been questioning simulated customers about topics such as pet medication and that CVS planned to scale its roster to 100,000 simulated people to test store layouts and new product designs [28].

### Series B (July 2026)

On 30 July 2026, five months after launch, Simile announced that it had raised "over $200 million at a $2 billion post-money valuation, co-led by Greenoaks and Index Ventures, with participation by Hanabi, Bain Capital Ventures, A*, Factory, CVS Health Ventures, and Definition" [2]. TechCrunch and Pulse 2.0 described [Greenoaks](https://aiwiki.ai/wiki/greenoaks_capital) as the lead investor, with Index participating [8][9], and The Next Web, citing The New York Times, also reported that Greenoaks led [27]. The company said that since launch it had grown revenue fivefold, had run "tens of millions of simulations for Fortune 100 enterprises", had trained a confidence model that predicts the accuracy of each simulation, and had expanded from its "small home in Palo Alto" to a global team of more than 50 employees [2]. It gave revenue growth only as a multiple, not as a figure [2]. Park told the newsletter Signals that the round was not planned: "We were preempted," he said, adding that investors "came to us and said, 'We'd love to invest now'" [11]. Bain Capital Ventures, which invested in both rounds, said it was "doubling down" [10].

By August 2026 Park said Simile had about 60 employees, a headquarters in San Francisco's Mission Rock area, where much of its technical staff is based, and a recently opened smaller office in New York [12].

## Research lineage

Simile's website presents the founders' academic papers as the basis of its products [1].

| Year | Work | Venue | Contribution |
|---|---|---|---|
| 2021 | "On the Opportunities and Risks of Foundation Models" | CRFM report (arXiv) | Introduced the term "foundation models"; Park, Bernstein and Liang were among the authors [18] |
| 2022 | "Social Simulacra: Creating Populated Prototypes for Social Computing Systems" | UIST 2022 | Language-model-generated communities for prototyping social platforms [15] |
| 2023 | "Generative Agents: Interactive Simulacra of Human Behavior" | UIST 2023 (best paper award) | Memory, reflection and planning architecture; 25-agent Smallville town [14][19] |
| 2024 | "Generative Agent Simulations of 1,000 People" (retitled in 2026) | arXiv preprint | Interview-grounded agents of 1,052 people; 85 percent of test-retest accuracy on the General Social Survey [16] |
| 2025 | "Finetuning LLMs for Human Behavior Prediction in Social Science Experiments" | EMNLP 2025 | SocSci210 dataset and fine-tuned "Socrates" models [17] |
| 2025 | "Generative Agent Simulations of Human Behavior" | Stanford PhD dissertation (Park) | Arthur Samuel Award [19] |

## Technology

### Data

Park describes three kinds of training data [12]. The first is long-form interviews, in which participants are asked to "tell me the story of your life". The second is observational behavioral data such as transactions. The third, which he considers perhaps the most important, is data on causal mechanisms from randomized controlled trials and other experiments in which participants face real stakes, for example an online store in which people actually receive what they buy [12]. Participants are recruited with consent and with incentives [12], and Park has said the company trains its own interviewer model in a [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) loop to learn as much about a person as possible in the least time [13]. Simile recruits people partly through a strategic partnership with the polling company Gallup [13]. A company research post says Simile trains on "transactions from grocery stores and delivery apps" and on "AI-led voice interviews and surveys on our in-house platform" [22].

In a September 2026 essay, Bernstein argued that behavioral logs show what people do but not why, that self-reports are distorted by a "say-do gap", and that a model of behavior therefore has to be trained on both [23]. Customers can also contribute their own data, such as loyalty records, which the company says is opt-in and trains a custom model for that customer [1].

### Simulation and confidence models

Simile describes its core technology as "a foundation model for human behavior" [1]. Park has said the company post-trains its own model and is seeing "the early glimpse of scaling law in simulations", with predictable gains as it adds human data and compute [12]. On Sequoia Capital's Training Data podcast he contrasted frontier models, which he called the "CPU of intelligence" because they are built for problems with right answers, with the "GPU of intelligence" that Simile wants to build to capture the diversity of human values, preferences and tastes [13].

An August 2026 technical post by Andrew Wesel, Sarah Chen and Percy Liang describes the interface: the model takes a natural-language prompt describing a population, a situation and a set of valid actions, and returns a probability distribution over those actions [22]. In the experiments reported in that post, the simulation model was a fine-tune of [Alibaba](https://aiwiki.ai/wiki/alibaba)'s open-weight [Qwen3.5](https://aiwiki.ai/wiki/qwen3_5)-27B [22]. Accuracy is scored by total variation distance (TVD) between the simulated distribution and one observed from real people. Simile set a "decision-grade" threshold of 0.16 TVD, chosen from about 2,750 ratings by 14 raters from its own team [22]; Park had earlier described a TVD below 0.15 as strong evidence for decision-making [13].

Because the true answer is unknown when a customer asks a new question, Simile trains a separate "confidence model" to predict each simulation's error. On about 8,600 held-out questions, the post reports an [area under the ROC curve](https://aiwiki.ai/wiki/auc_area_under_the_roc_curve) (for separating simulations above and below the 0.16 threshold) of 0.57 for a regression on basic features, 0.69 using a general-purpose text embedding, 0.73 for a probe on the simulation model's hidden states and 0.74 for continued fine-tuning of the simulation model itself [22]. The product sorts results into "High", "Moderately High", "Medium" and "Low" confidence; according to the post, a "High" result is 95 percent likely to meet the 0.16 threshold, against 38 percent for "Low" [22]. Simile also says it validates against real people every week, with more than 7,000 evaluations, and refreshes data when it detects drift [1].

In a March 2026 post, Liang called simulation "the next frontier of AI". He argued that a simulator is a causal model that can answer interventional and counterfactual questions in the sense of [Judea Pearl](https://aiwiki.ai/wiki/judea_pearl)'s causal hierarchy, and described simulation as "the true path to robust superintelligence" [24].

## Products and customers

A customer describes a population, such as people in their 20s and 30s living in California, and an "environment", which can be survey questions, behavioral experiments or [A/B tests](https://aiwiki.ai/wiki/a_b_testing). Simile recruits and models that population, and the customer then queries it through the company's software [12]. Uses Park has named include concept testing, focus groups, product testing in which agents browse websites or design mockups, and multi-agent simulations of public companies' earnings calls [12][13]. According to The Next Web, Simile prices partly by the number of agents it models, from hundreds of thousands to millions [27]. Bain Capital Ventures describes each "simile" as mapping one-to-one to a real individual [10].

| Customer | Reported use | Sources |
|---|---|---|
| CVS Health | Validating known insights, customer-satisfaction drivers, medication adherence and competitive positioning | [25][6] |
| Wealthfront | Early customer for product testing in which agents use websites and mockups; its head of user research said Simile expanded the scope of its qualitative research "by 15X" | [12][1] |
| Itaú Unibanco | Tested product hypotheses for Pix Automático, the recurring-payments layer of Brazil's instant payment system | [26] |
| Gallup | Strategic partnership | [13][12] |
| Deloitte | Named by the company as a customer | [2] |
| Suntory Beverages & Food | Product development partnership (company testimonial) | [1] |
| Telstra Group | Named by SiliconANGLE as an early customer | [7] |
| Garnett Station Partners | Consumer research for a private equity team (company testimonial) | [1] |

Simile's February 2026 post on CVS Health says the agents used there are built on 2.9 million consented responses from more than 400,000 participants across more than 200 behavioral scenarios. It lists faster validation of known insights, sharper drivers of customer satisfaction, tests of refill reminders and benefit designs, and competitive benchmarking as results [25]. Those figures are close to the size of the SocSci210 research dataset described above, though the post does not say whether the same data was used [25][17]. For Itaú Unibanco, Simile says a research cycle that normally took five weeks was completed in four business days, and that the bank's paper on the work was named the Latin America regional winner for Excellence in AI and Automation in Market Research at the 2026 ESOMAR Congress in Valencia [26].

## Funding

| Round | Announced | Amount | Lead investors | Other named investors | Valuation |
|---|---|---|---|---|---|
| Series A | 12 February 2026 | $100 million [3][5] | [Index Ventures](https://aiwiki.ai/wiki/index_ventures) [3] | Hanabi, A*, Bain Capital Ventures; individuals [Andrej Karpathy](https://aiwiki.ai/wiki/andrej_karpathy), [Fei-Fei Li](https://aiwiki.ai/wiki/fei_fei_li), Adam D'Angelo, Guillermo Rauch and Scott Belsky [3] | Not disclosed [6] |
| Series B | 30 July 2026 | More than $200 million [2] | [Greenoaks](https://aiwiki.ai/wiki/greenoaks_capital) and Index Ventures (company) [2]; Greenoaks (TechCrunch, Pulse 2.0) [8][9] | Hanabi, Bain Capital Ventures, A*, Factory, CVS Health Ventures, Definition [2] | $2 billion post-money [2] |

## People

| Name | Role at Simile | Background |
|---|---|---|
| Joon Sung Park | Co-founder and chief executive [1] | BA in computer science from Swarthmore College, MS from the University of Illinois Urbana-Champaign, PhD in computer science from Stanford; first author of Social Simulacra, Generative Agents and the 1,000-person study; previously worked as an oil painter and held a solo exhibition in 2017 [19][4] |
| Michael Bernstein | Co-founder and chief data officer [1] | Professor of computer science at Stanford and senior fellow at the [Stanford Institute for Human-Centered Artificial Intelligence](https://aiwiki.ai/wiki/stanford_hai); BS from Stanford, PhD from MIT; Sloan Research Fellowship and the Computer History Museum's Tech for Humanity Prize [20] |
| Percy Liang | Co-founder and chief scientist [1] | Stanford computer science professor and director of CRFM; creator of [SQuAD](https://aiwiki.ai/wiki/squad) and [HELM](https://aiwiki.ai/wiki/helm), according to Simile; Presidential Early Career Award for Scientists and Engineers (2019) and IJCAI Computers and Thought Award (2016) [21][1] |
| Lainie Yallen | Co-founder; Park's business counterpart [12] | Helped scale companies including [Hebbia](https://aiwiki.ai/wiki/hebbia), Valence and TriplePlay before joining Simile, according to Bain Capital Ventures [10] |

Park has said that while "there are AI neo labs and then there are AI product companies", Simile "truly is both", with three researcher founders and a fourth who runs the commercial side, and that the company has hired some of his former lab mates [12].

## Reception and criticism

TechCrunch's Julie Bort called the stated mission of simulating "all eight billion people on earth" "preposterous", arguing that "the whole reason to conduct market research is because humans are unpredictable", while calling simulated users for research "a promising area" and comparing it to vibe coding for product mock-ups [8]. She placed Simile alongside Aaru, which TechCrunch said had raised a Series A the previous December at a "headline" $1 billion valuation [8]. The Next Web noted that Simile's claims of 85 to 99 percent accuracy, depending on the group, come from the company itself, called the premise "contested", and named Helm, Artificial Societies and Aaru as rivals [27]. Gizmodo, covering the Wall Street Journal profile, compared the product to The Sims and answered a CVS finding about pet medication with "Huge if true" [28].

The broader idea of using language models as stand-ins for survey respondents predates Simile. Argyle and colleagues proposed "silicon samples" in 2022, arguing that a language model conditioned on real participants' backstories can reproduce the response patterns of demographic subgroups [30]. Wang, Morgenstern and Dickerson, in a paper accepted at Nature Machine Intelligence, argued that language models used to replace human participants tend to "misportray and flatten" demographic groups, and urged caution where participants' identities matter [31]. That work does not evaluate Simile's system. The founders' 2024 study addressed part of this concern by grounding agents in individual interviews rather than demographic prompts, and reported smaller accuracy gaps across racial and ideological groups than demographics-only agents [16].

Simile appears on Debarghya Das's "The Ultimate Neolab List", which labels it "Borderline" and explains: "Researcher-founded and raised big early, but product-led with enterprise customers and modest compute." [32]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [AI agents](https://aiwiki.ai/wiki/ai_agents)
- [Multi-agent system](https://aiwiki.ai/wiki/multi_agent_system)
- [Agent memory](https://aiwiki.ai/wiki/agent_memory)
- [Digital twin](https://aiwiki.ai/wiki/digital_twin)
- [Synthetic data](https://aiwiki.ai/wiki/synthetic_data)
- [Percy Liang](https://aiwiki.ai/wiki/percy_liang)

## References

1. Simile, [Simile homepage](https://www.simile.com/). Accessed 27 October 2026.
2. The Simile Team, ["Announcing Simile's Series B"](https://www.simile.com/blog/series-b), Simile blog, 30 July 2026.
3. The Simile Team, ["The Simulation Company"](https://www.simile.com/blog/the-simulation-company), Simile blog, 12 February 2026.
4. Shardul Shah, ["Life, the Universe, and Simile: Leading Simile's Series A"](https://www.indexventures.com/perspectives/life-the-universe-and-simile-leading-similes-series-a/), Index Ventures, 12 February 2026.
5. Latham & Watkins, ["Latham & Watkins Advises Simile AI, Inc. on US$100 Million Series A"](https://www.lw.com/en/news/latham-advises-simile-ai-inc-on-us-100-million-series-a), 19 February 2026.
6. Bloomberg, ["Startup building model to predict human behavior"](https://www.taipeitimes.com/News/biz/archives/2026/02/15/2003852338), syndicated in the Taipei Times, 15 February 2026.
7. Maria Deutscher, ["AI digital twin startup Simile raises $100M in funding"](https://siliconangle.com/2026/02/12/ai-digital-twin-startup-simile-raises-100M-funding/), SiliconANGLE, 12 February 2026.
8. Julie Bort, ["Synthetic-user startup Simile raises $200M at $2B valuation 5 months after $100M Series A"](https://techcrunch.com/2026/07/30/synthetic-user-startup-simile-raises-200m-at-2b-valuation-5-months-after-100m-series-a/), TechCrunch, 30 July 2026.
9. Amit Chowdhry, ["Simile Raises Over $200 Million In Series B At A $2 Billion Valuation"](https://pulse2.com/simile-raises-over-200-million-in-series-b-at-a-2-billion-valuation/), Pulse 2.0, 30 July 2026.
10. Bain Capital Ventures, ["The Human Layer of AI: Why we're continuing to invest in Simile"](https://baincapitalventures.com/insight/the-human-layer-of-ai-why-we-re-continuing-to-invest-in-simile/), 30 July 2026.
11. Kate Park, ["On the Record #9: Why Simile raised $200M in five months without planning to, reaching a $2B valuation"](https://signalsventures.substack.com/p/on-the-record-9-why-simile-raised), Signals: Inside the Story, 14 August 2026.
12. Latent Space, ["Simulation: the new Scaling Law"](https://www.latent.space/p/simile) (podcast with Joon Sung Park, with transcript), 21 August 2026.
13. Sequoia Capital, ["Simulating Humans at Scale: Simile's Joon Sung Park"](https://sequoiacap.com/podcast/simulating-humans-at-scale-similes-joon-sung-park), Training Data podcast (with transcript). Accessed 27 October 2026.
14. Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang and Michael S. Bernstein, ["Generative Agents: Interactive Simulacra of Human Behavior"](https://arxiv.org/abs/2304.03442), arXiv:2304.03442, April 2023 (UIST 2023).
15. Joon Sung Park, Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Percy Liang and Michael S. Bernstein, ["Social Simulacra: Creating Populated Prototypes for Social Computing Systems"](https://arxiv.org/abs/2208.04024), arXiv:2208.04024, August 2022 (UIST 2022).
16. Joon Sung Park et al., ["Generative Agent Simulations of 1,000 People"](https://arxiv.org/abs/2411.10109v1), arXiv:2411.10109v1, 15 November 2024; revised as ["LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals"](https://arxiv.org/abs/2411.10109), v3, 28 June 2026.
17. Akaash Kolluri, Shengguang Wu, Joon Sung Park and Michael S. Bernstein, ["Finetuning LLMs for Human Behavior Prediction in Social Science Experiments"](https://arxiv.org/abs/2509.05830), arXiv:2509.05830; [Proceedings of EMNLP 2025](https://aclanthology.org/2025.emnlp-main.1530/), pages 30096-30111.
18. Rishi Bommasani et al., ["On the Opportunities and Risks of Foundation Models"](https://arxiv.org/abs/2108.07258), arXiv:2108.07258, Stanford Center for Research on Foundation Models, August 2021.
19. Joon Sung Park, [personal website](https://www.joonsungpark.com/) (biography, publications and exhibitions). Accessed 27 October 2026.
20. Stanford Profiles, ["Michael Bernstein"](https://profiles.stanford.edu/michael-bernstein?tab=bio). Accessed 27 October 2026.
21. Stanford Profiles, ["Percy Liang"](https://profiles.stanford.edu/percy-liang?tab=bio). Accessed 27 October 2026.
22. Andrew Wesel, Sarah Chen and Percy Liang, ["Building confidence in Simile"](https://www.simile.com/blog/confidence), Simile blog, 25 August 2026.
23. Michael Bernstein, ["Building the What-If Machine"](https://www.simile.com/blog/what-if-machine), Simile blog, 2 September 2026.
24. Percy Liang, ["Simulation: The Next Frontier for AI"](https://www.simile.com/blog/simulation-next-frontier), Simile blog, 10 March 2026.
25. Simile Business Staff, ["CVS Health x Simile: Simulations for faster, safer decisions"](https://www.simile.com/blog/simile-cvs-health), Simile blog, 24 February 2026.
26. Simile Business Staff, ["How Itaú turned five weeks of research into four days"](https://www.simile.com/blog/itau-at-esomar), Simile blog, 11 September 2026.
27. Cristian Dina, ["To find out what customers think, this $2bn startup asks their AI twins"](https://thenextweb.com/news/simile-200-million-agentic-twins-ai-market-research), The Next Web, 30 July 2026.
28. Mike Pearl, ["An AI Company Apparently Inspired by 'the Sims' Wants to Revolutionize Public Opinion Research"](https://gizmodo.com/an-ai-company-apparently-inspired-by-the-sims-wants-to-revolutionize-public-opinion-research-2000731038), Gizmodo, 8 March 2026.
29. joonspk-research, ["generative_agents"](https://github.com/joonspk-research/generative_agents) (code repository, Apache 2.0), GitHub. Accessed 27 October 2026.
30. Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua Gubler, Christopher Rytting and David Wingate, ["Out of One, Many: Using Language Models to Simulate Human Samples"](https://arxiv.org/abs/2209.06899), arXiv:2209.06899, September 2022.
31. Angelina Wang, Jamie Morgenstern and John P. Dickerson, ["Large language models that replace human participants can harmfully misportray and flatten identity groups"](https://arxiv.org/abs/2402.01908), arXiv:2402.01908 (accepted at Nature Machine Intelligence).
32. Debarghya Das, ["The Ultimate Neolab List"](https://debarghyadas.com/neolabs/). Accessed 27 October 2026.
33. Dealroom, ["Simile"](https://dealroom.co/companies/simile/) (company profile). Accessed 27 October 2026.