Simile
In 2023, Joon Park put 25 AI agents in a simulated town. They planned a party on their own.
Verdict: Simile's AI twins helped CVS study whether patients take their medicine, but its homepage sells launch and price testing.
A Narrative Authenticity Framework analysis of Simile. NAF score 8.4 out of 10. Weakest component: Real Aspiration, 7.0.
Originally published on LinkedIn, May 1, 2026. Facts checked and updated November 2026.
In 2023, Stanford PhD student Joon Sung Park and his co-authors put 25 AI agents in a simulated town called Smallville. Nobody scripted them. They woke up, went to work, spread gossip and organized a Valentine's Day party on their own.
Park turned that research into Simile with two of his co-authors, Stanford professors Michael Bernstein and Percy Liang. The company left stealth in February 2026 with $100 million. Five months later it raised $200 million more at a $2 billion valuation.
One of its earliest uses had nothing to do with shelf space. The Next Web reported that CVS Health used 400,000 of Simile's AI twins to study how to get patients to take their medicine.
That is the most human thing the company has done. Its homepage talks about launches, prices and campaigns.
Where Simile stands
The founders. CEO Joon Sung Park, Chief Product Officer Michael Bernstein, Chief Scientist Percy Liang, who helped coin the term "foundation model," and Chief Customer Officer Elaina Yallen.
The money. $100 million in February 2026, led by Index Ventures. $200 million in July 2026 at a $2 billion valuation, led by Greenoaks, per the New York Times.
The customers. The homepage names CVS, Gallup, Itaú, Deloitte, Wealthfront and Suntory.
The brand line. "Simulate every decision at scale," with the promise to show "how and why customers, employees, or populations respond to change."
The NAF scorecard
| Component | Score | One-line read |
|---|---|---|
| Friction | 8.5 | "Launch and hope" for decisions that affect millions |
| Lived Stake | 8.5 | The researchers who started the field |
| Distinctive Qualification | 9.5 | Peer-reviewed research, top investors, real customers |
| Real Aspiration | 7.0 | A bigger world in the founders' words than on the site |
| Composite | 8.4 |
Scores as of publication. Current standings at seekargus.com/leaderboard.
The scores hold from the original article. Simile shares 8.4 with GitHub and Patreon.
Friction: 8.5
Bernstein named the problem on launch day: when organizations make decisions that affect billions of people, they often "launch and hope." Focus groups are slow and small, and real launches are expensive ways to learn.
It stops short of 9 because rivals such as Aaru, Artificial Societies and Helm pitch the same cost and speed problem. There's also a second friction Simile hasn't named: buyers need to trust answers that come from synthetic people.
Lived Stake: 8.5
The founders didn't spot a market. They did the research that created it, through peer review, before building a company.
It's an intellectual stake more than a personal one. None of them lived "launch and hope" as an operator, which keeps the score below 9.
Distinctive Qualification: 9.5
The Generative Agents paper introduced the memory, planning and reflection loop that many agent systems now use, and a follow-up study built agents of 1,000 real people from two-hour interviews. Investors, including some of the best-known names in AI, have now put $300 million behind it.
One caution belongs in the record. Simile's 85 to 99 percent accuracy figures are its own, as The Next Web noted, and critics question whether AI stand-ins capture how people actually behave.
Real Aspiration: 7.0
The founders speak in large terms. Park told the New York Times that simulating the world lets you "test out countless interventions" and called it "a window into humanity." The launch materials talked about simulating entire worlds.
The homepage promises to simulate decisions at scale, which describes a capability. A 9 would describe the changed world: no consequential decision reaches real people until it has been tested on simulated ones, starting with health.
The story Simile should tell
In 2023, we put 25 AI agents in a small town and watched them plan a party nobody told them to plan. Today we use the same science to help a pharmacy learn how to get patients to take their medicine. Every decision that touches millions of people should be tested before it reaches them.
Three moves to raise the score
1. Put the changed world on the homepage (Real Aspiration 7.0 to 8.0)
Keep "Simulate every decision at scale" as the capability, and add one line on what it changes: decisions get tested on simulated people before they touch real ones. Measure it by how often press and buyers repeat the line.
Prior-work check: Partial. The homepage line is broader than the old "simulation platform for human behavior," and Park speaks about testing interventions in interviews. The homepage doesn't state a changed world, as of November 19, 2026.
2. Publish how the twins are built and consented (Friction 8.5 to 9.0)
Explain, on one public page, how interview participants agreed to have AI twins built from their lives, what they're told, and how Simile checks twins against real people. Measure it by procurement cycle time and by fewer trust objections in sales.
Prior-work check: Partial. The homepage says customer data is "opt-in and customer-governed" and that results are validated "against real humans weekly." No public page explains how interview participants consented, and The Next Web's July 2026 coverage didn't address it, as of November 19, 2026.
3. Make the CVS medication study the flagship case (Real Aspiration 8.0 to 8.5)
Publish the medication study, with CVS's approval, as the proof of where Simile is headed: health decisions tested before they reach patients. Measure it by inbound interest from health and public-sector buyers.
Prior-work check: Partial. CVS appears on the homepage, and the medication work was reported by the press in July 2026. No Simile case study on it found, as of November 19, 2026.
| Move | Closest existing program | Status |
|---|---|---|
| Changed world on the homepage | "Simulate every decision at scale" | Partial |
| How twins are built and consented | Homepage data and validation lines | Partial |
| CVS medication study as flagship | CVS logo; press coverage | Partial |
Projected composite: 8.9 (Friction 9.0, Lived Stake 8.5, Distinctive Qualification 9.5, Real Aspiration 8.5). These checks cover public sources only and can't see Simile's internal plans.
Sources
- Simile homepage, viewed November 19, 2026
- The Next Web: Simile raises $200 million at a $2 billion valuation, July 30, 2026
- New York Times, via The Star: Park on "a window into humanity", July 2026
- Dealroom: Simile founders and February 2026 round
- Implicator: the $100 million Series A, 2026
- Park et al., "Generative Agents: Interactive Simulacra of Human Behavior", 2023
Tim Donovan · SeekArgus