Company teardown · Simile

How to simulate a human being?

Simile trains foundation models that behave like people and sells them to enterprises in place of the research panel. Its founders: two Stanford professors and the doctoral student who recruited them. “Company is a machine for depth research.” $300M and fifty-nine hires later, here's who they hired.

Get the full visual brief as a PDF.
Founded
2025 · Palo Alto
People
59 of ~60 staff
Raised
$300M · $2B valuation
Customers
Gallup · CVS · Telstra
Sells
Simulated customers

0.1 — Simile in one page

A Stanford paper, turned into a company — then sold by the field force of the category it replaces.

59
People in fifteen months
Twelve before the money, forty-seven after — a nucleus in 2025, an organization in 2026.
$300M
Raised, at a $2B valuation
$100M in February from Index, $200M by July from Greenoaks — fifteen months after incorporation.
26: 23
Sell and deploy vs build
More sellers and deployers than builders; more MBAs than PhDs; six people with a publication record.
1 · The idea
Park's 2023 paper gave language models a memory and a town to live in; his 2024 study matched 1,052 real people 85% as well as they matched themselves. Simile sells that: a model that answers the way a person would.
2 · The business
Synthetic users and digital twins for enterprises that used to buy research panels — Gallup, CVS, Suntory, Telstra. $300M raised in 2026 at a $2B valuation; sold by an enterprise field force, delivered by ex-consultants.
3 · The founders
Joon Sung Park recruited his own PhD advisors — Michael Bernstein and Percy Liang — as co-founders, and Lainie Yallen to run operations. The seniority runs the wrong way, on purpose.
4 · The team
Fifty-nine people, forty-seven of them hired in 2026. More sellers and deployers (26) than builders (23); more MBAs than PhDs; six with a publication record. Three sites, three different shapes.
5 · The lessons
Recruit upward, land an anchor and let the cohort form, buy the incumbent's field force, hire the people who built the benchmarks — and know that cash won't close the researchers.

Everything here is public — LinkedIn, funding disclosures, three founder appearances with transcripts, and a company film. The reading is ours.

§ 01 — Part one of five · The idea

What if you could ask a thousand people before you asked anyone?

Simile's founders spent five years teaching language models to behave like specific people. The company sells the result to anyone who used to buy a research panel.

1.1 — The problem they're solving

Why every decision about people still waits on a panel.

Reason oneThe panel is slow and dearA survey “might take months and potentially millions of dollars.” By Park's count, 5% of a company's questions get tested — “the rest of the 95% we never bother.”
Reason twoYou can't ask the counterfactual“No one really cares about prediction… they want to shape the future.” A respondent can't say what they'd do at another price; a simulated one can be run again.
Reason threePeople are noisyRe-asked the same questions two weeks later, participants in Park's 1,000-person study matched their own answers about four times in five. Human consistency is the ceiling.

— The founder's framing

“The only way to find out whether your decision was correct was by living through it.”

— Joon Sung Park, Simile “simulation era” film, 2026.

— Behavioral simulation, translated

A model trained to predict what a specific person — or a population — would decide. Park's analogy: the frontier LLM is “the CPU of intelligence”; Simile builds “the GPU” — many units, each as flawed as the person it stands for.

1.2 — The idea

A survey respondent you can run again.

Interview real people once; give a model the interview as memory; then ask it anything, as often as you need. The test is whether it answers the way the person would.

01 · Capture

Interviews and behavior “Tell me the story of your life”: a two-hour interview per person, plus “a huge repo of randomized control trials” — pricing, behavior.

02 · Simulate

Foundation models of people “We want our models to make the same kind of mistake” — biased the way people are. Population models and digital twins, sold in place of a panel.

03 · Check

Agreement with the humans Total variation distance from the real answers; under 0.15 is “strong evidence for making a decision.” Diverging runs repeat for confidence.

↺ “the world is our ground truth” — hypotheses graded monthly

Generative agentA model with a life storyA language model given one person's memories and asked to act as they would. Park's 2023 paper put 25 of them in a sandbox town, Smallville; they planned their days, spread news and organised a party unprompted.
Conformal predictionHonest confidence intervalsStatistics for attaching an honest confidence interval to a model's output. Park's diverging simulations — “run it 100 times; how many come out X?” — need exactly this: the specialty Simile hired for in August.

1.3 — The lineage

Twenty-five agents, then a thousand people, then the economy.

— What's Smallville?

The 2023 Generative Agents paper: 25 language-model agents in a pixel-art town, each with memory, reflection and plans. Human evaluators rated their behaviour believable; the architecture has been copied across the field since. 8,125 citations — the paper Simile productizes.

— Five years, four steps

2021–22 · The groundwork
Park's first Stanford year goes into the report that named “foundation models”; then Social Simulacra — a subreddit simulated before it exists.
2023 · Generative Agents
Smallville. Memory, reflection, planning — the architecture that made language models act like people.
2024 · 1,000 people
1,052 real Americans interviewed for two hours each; their agents matched their survey answers 85% as well as they matched themselves two weeks later.
2025 · The company
Population- and individual-level foundation models, incorporated mid-2025. The 20VC episode title a year later: “The AI Company Simulating the Entire Economy.”

The academic bench is inside the company: Michael Bernstein (Stanford HCI) and Percy Liang (Stanford CRFM, who coined “foundation model”) are co-founders; Fei-Fei Li is an angel.

§ 02 — Part two of five · The business

A research lab with a sales floor.

Simile isn't selling a paper. It sells simulated customers to enterprises — through a field force hired from the companies that used to sell them surveys.

2.1 — The product, as described

Synthetic customers for the enterprise — sold one deployment at a time.

01
Who buys
Enterprises that used to buy panels
CVS, Wealthfront, Suntory, Telstra, Deloitte, Itaú Unibanco — and Gallup, customer and “strategic partner” that reaches the real people each model is grounded in. The incumbents on both sides of the product.
02
What they get
Synthetic users · digital twins
Concept tests — “a thousand ideas across a thousand sub-populations” — and multi-agent runs: “some of our customers routinely ask us to simulate their earnings call.”
03
How it's sold
Enterprise sales, then Deployments
A field force closes — the largest customers “close deals within 3 months” — and a “Deployments” team delivers, staffed from Bain, BCG and Palantir. Thirteen carry the title; most sit in New York.
04
What it isn't
A developer platform
No GitHub org, no Hacker News thread, no public docs; the careers page needs a login. A $2B AI company with almost no developer surface — by choice.
05
Today
Case studies and awards
The Itaú Unibanco study won Esomar's Latin America award for AI in market research (Sep 2026); Bernstein hosts a public technical session on simulation data on the 24th.

Deployments: Simile's title for the forward-deployed role — the Palantir motion, run by consultants. One Deployments hire is a former Palantir forward-deployed engineer.

2.2 — The money

$100M in February, $200M by July.

$100M
Series A — Index Ventures, February 2026.
$200M
Series B — Greenoaks, closed 31 July 2026.
$2B
Post-money — fifteen months after incorporation.
2–10
Employees, per the company's own LinkedIn band — against 67 associated members.
I1
Index Ventures · Greenoaks
The leads · Series A · Series B
Index led the A; insiders pre-empted the B and Greenoaks priced it at $2B. The seed: Mike Volpi and A* — Volpi introduced Park to his COO.
I2
Hanabi · Bain Capital Ventures · A*
The syndicate · Factory · Definition
Bain Capital Ventures on the cap table; Bain & Company ×3 on the Deployments team; Bain Capital in one deployer's résumé. One alumni pool.
I3
CVS Health Ventures
Customer & investor · Series B participant
A customer, a Series B investor and a feeder — a product lead into Deployments, a CX executive into GTM. The buyer sits on all three sides of the table.
I4
Fei-Fei Li · Andrej Karpathy
The angels · personal cheques
Both share a byline with Bernstein on the ImageNet challenge paper — 56,960 citations, half his total. The investors are the founders' bibliography.

2.3 — Who else is in the room

The incumbents aren't the rivals — they're the feeder.

01
Medallia
The category replaced · experience management · survey research
Sells the surveys and CX research that synthetic panels are pitched against. Five Medallia alumni now sell for Simile — every one hired in 2026.
02
Cresta
The adjacent incumbent · conversational AI for contact centres
Six alumni — the roster's second-largest feeder — beginning with the Founding Global Sales Director in February 2026, the month of the Series A.
03
Gallup
The panel, as supplier · polling house · customer and partner
A “strategic partnership with Gallup” reaches the real people each simulation is grounded in. The most famous name in survey research supplies the replacement.
04
Deloitte · BCG · Bain · McKinsey
The consultancies · one customer, three feeders
The firms that used to run the research projects for clients now staff “Deployments”: BCG ×4, Bain ×3, McKinsey ×2, eight MBAs.
05
Nothing
The public square · 30-day scan
“Simulation” and “digital twin” collide with gaming vocabulary; the company query found no thread anywhere on X, Reddit, Hacker News or GitHub. No competitor set could be built from public signal.

For a recruiter this is one market — and Simile hires the people who sold the incumbent product to the same buyers.

Talent brief

Want the full visual breakdown?

Download the PDF version of this teardown — the six-team org chart, ten names to know, the build order month by month, and the six sourcing patterns none of which ran through a job board.

2.4 — The business, in his words · 20VC · Sequoia Training Data

Six things Park says about the business — and what each explains on the roster.

01 · The modelWrong the way people are wrong“We want our models to make the same kind of mistake… biased in the same way humans are.” Frontier labs build “the CPU of intelligence”; Simile the GPU — many units, none superhuman. The bench is HCI and social computing, not infrastructure.
02 · The dataPrediction is overrated“No one really cares about prediction… they want to shape the future.” So the training asset is causal: “a huge repo of RCTs,” pricing studies, A/B tests. A Head of Data was hired in the first month.
03 · The flywheelThe world is our ground truth“Every single day we can be generating tens of thousands of hypotheses… a month goes by, x percent came true.” The reward coding agents get from accept/reject, simulation gets from the calendar. Hence an Evals bench of four.
04 · The metricTVD under 0.15Population answers are scored by total variation distance from the real distribution; “less than 0.15 is strong evidence for making a decision.” Diverging runs repeat a hundred times for confidence — the conformal-prediction hire, in August.
05 · The wedgeAsk people to pay youPat Hanrahan's advice, kept: “the best way to get feedback is to ask people to pay you.” Enterprise research is the wedge — “deals within 3 months,” six-month studies rerun “within 2 minutes.” Twenty-six people sell and deploy it.
06 · The horizonA $100M simulation“In about two to three years we're running a single simulation session that people will pay $100 million for.” Synthetic panels “larger than the current human panel market.” The vision is the comp — what he sells researchers instead of base salary.

§ 03 — Part three of five · The founders

Two professors, their student — and the student is the CEO.

Four co-founders: three Stanford researchers and one operator. The seniority runs the wrong way, on purpose.

3.1 — Who they are

Four co-founders, one lab and one operator.

— Joon Sung Park · Co-founder & CEO · the student

Swarthmore · UIUC
A figure painter and CS major with no undergraduate research; master's at Illinois.
Stanford
PhD under Bernstein and Liang; first author of Generative Agents, UIST 2023.
Awards
Best papers at UIST and CHI; Microsoft Research PhD Fellowship; Siebel Scholar.
Simile
Incorporated mid-2025; recruited both advisors as co-founders. 25,119 citations.

— Michael Bernstein · Co-founder & Chief Data Officer · the advisor

Stanford · MIT
Symbolic Systems BS; MIT PhD — Soylent, the crowd-powered word processor.
ImageNet
Co-author of the ImageNet challenge paper (56,960 citations) and Visual Genome.
Stanford HCI
Professor; seven best-paper awards, NSF CAREER, Sloan Fellowship. Park's advisor.
Simile
Joined July 2025. 116,460 citations, h-index 87 — the roster's largest by far.
Percy Liang · Co-founder & Chief ScientistThe man who named the categoryStanford professor; founding director of CRFM — coined “foundation model.” Absent from the LinkedIn page, so every roster ratio here undercounts the research bench by one.
Lainie Yallen · Co-founder & COOThe operatorMcGill, BCG, Hebbia. Joined July 2025; runs the enterprise side from New York — “I was not an enterprise seller. I needed Laney to lead that part of the game.”

3.2 — What the CEO says · Latent Space · 20VC · Sequoia Training Data

Six things Park says — and what the roster says back.

01 · The labThe lab is the pitch“15%, almost 20% of the company population are just my lab mates.” Nine of 59 carry a working Stanford affiliation: exactly 15%. The transcript says “Microsoft Research”; the roster says Stanford.
02 · The screenTwo superpowers that shouldn't coexist“Any expert will usually come in with one superpower … but they're all correlated.” He hires for uncorrelated pairs. The roster's hybrids: a Schwarzman Scholar in Deployments, a Citadel quant in engineering, an Evals researcher from Twitch.
03 · The proofThe common denominator of success“Are we the common denominator of success?” — did they win in several places, or one lucky one? The roster is built of second and third acts: Figma after Meta, Bain after PwC, OpenAI after Reddit.
04 · AcademicsMarried to impact, not a problemHis test for recruiting researchers: “are they married to a problem or married to impact?” Six said yes — including a Stanford PhD student who joined mid-degree.
05 · CompCash doesn't close themPeers' “total comp does range in tens of millions … meeting them at their base salary is tricky.” He sells vision instead. The bench is six publishing researchers of 59.
06 · The businessA machine for depth research“Research is an amazing vehicle if you want to do breadth research … company is a machine for depth research.” The company's depth: one paper, one product, twenty-six people selling and deploying it.

3.3 — How he builds a team · 20VC

The team as self-portrait.

— The principle

“Your subject looks like the painter. You should see yourself in the team.”

— Joon Sung Park, a figure painter before he was a researcher. 20VC, August 2026.

— How to read the table

Four principles, each one a thing Park says about building the team — set against what the roster actually shows. The commercial half of this company is designed, not drift.

Balance
“I was a researcher. I was not an enterprise seller. I needed Laney to be my co-founder to lead that part of the game.” The roster: 26 of 59 sell or deploy, run from New York by the COO the seed investor introduced.
Gaps, ahead of time
“As we scale there are new gaps that are emerging — seeing that ahead of time and making sure that we fill those gaps.” The roster: three recruiters at four-month intervals; an Evals bench before the Series B, a statistician the month after.
Reinvention
“Michael… started in crowdsourcing, then AI, now generative agents and simulations. At each step you could see: this is very Michael.” The roster: the common-denominator test, run on his own advisor. Second acts throughout — Figma after Meta, Bain after PwC, OpenAI after Reddit.
Paranoid, religious
“Laney… she's paranoid: unless we put everything on the table today, we'll lose. But long term, she's religious.” The roster: the operator's archetype. The sales floor opened the month the A closed — ten months ahead of Park's own plan to “go aggressive.”

3.4 — How they hire

“Places where we have personal connections” — tested against the roster.

Told to Latent Space: engineers come from Figma, Notion and Rive; lab mates follow the lab; “we're hiring across all sections.” Yet LinkedIn lists no open roles and the careers page sits behind a login.

— The claims, in the founder's words

“Figma, Notion, Rive”
Figma 7 — the top feeder · Notion 1 · Rive 0, at 88% coverage.
“15%, almost 20%”
9 of 59 with a working Stanford tie — 15% exactly.
“OpenAI, Google Gemini”
Kairam (OpenAI) · Kulkarni (Gemini 2.5 contributor) · Skylar Wang (DeepMind) — three, not a feeder.
“We do have quants”
Citadel · Citadel Securities · DRW · PIMCO · D. E. Shaw · ADIA · J.P. Morgan.

— How the channel is run

Three recruiters
Tommy F (Dec 2025, Scale AI) · Guanzing (May, Scale AI · Lambda) · Hakimjavadi (Sep, Ironclad · HashiCorp).
Zero postings
No roles on LinkedIn; careers.simile.ai requires a login.
Anchor, then cohort
Head of Product from Figma in Nov 2025; six more Figma alumni Feb – Jul 2026.
Buy the incumbent's field
Cresta 6 · Medallia 5 — eight sellers, all hired in 2026.

Three claims confirmed, one partial; the retention claim (“nobody left in six years”) waits for 2027. His account is accurate — and the “part of the game” he handed to Yallen is now the larger half of the company.

§ 04 — Part four of five · The team

Fifty-nine people: twelve before the money, forty-seven after.

The founders' claims, turned into an org chart — and where the research story and the roster part ways.

4.1 — The org chart, from the bets

Six teams — each one a founder's claim, staffed.

T1
Model · research
The bet: the next scaling law
Park · Bernstein · Kulkarni · Tony Lee · Ding · Kairam. Stanford HCI and CRFM, Microsoft AI, OpenAI, a Berkeley statistician. Six of 59 publish.
T2
Model · engineering
The bet: personal connections
Pondoc · White · Chen · Rathod · Chouhan · Ramos · Zhen. Figma ×3, SAIL, Glean and BAIR, Discord, Citadel: product engineers, not infrastructure.
T3
Evaluation
The bet: accuracy is the product
Tony Lee · Kairam · Srinivasan · Ding. HELM's builder, two “Evals” titles, conformal prediction — the people who built the field's measuring instruments.
T4
Product & design
The bet: ship it like a design tool
Kapoor · Stoy · Catania · Keeyen Y. · Tenggoro. Figma ×4, Notion, Reve, Twitter. The Head of Product was hired in month five.
T5
Deployments
The bet: consultants deliver it
Fox · Quill · Moonjely · Salzman · Boyd · Musuvathy. Bain, BCG, Palantir, CVS, Stanford HAI: thirteen titles, centred on New York.
T6
Sales & GTM
The bet: buy the field force
Jensen · Wymer · Brown · O'Hara · Hittson · Aronow. Cresta, Medallia, McKinsey, D. E. Shaw. Thirteen people; seven work from cities with no office.

4.2 — Ten names to know

The hires that set the bar.

01
Joon Sung Park
Co-Founder & CEO
Stanford CS PhD; first author of Generative Agents (8,125 citations) and the 1,000-person study.
02
Michael Bernstein
Co-Founder · Chief Data Officer
Stanford professor, MIT PhD. ImageNet challenge and Visual Genome co-author. 116,460 citations, h-index 87.
03
Lainie Yallen
Co-Founder & COO
McGill, BCG, Hebbia. The operator among the founders; runs the company from New York, where deployments and product sit.
04
Tony Lee
MTS · from Stanford SAIL
Built HELM (4,106 citations); VHELM first author; co-author on Llama 4. Stanford PhD in progress. Joined Jul 2026.
05
Chinmay Kulkarni
Member of Technical Staff
Stanford PhD; faculty at CMU and Emory; a year at Microsoft AI; Gemini 2.5 contributor. h-index 35. Joined Jun 2026.
06
Tiffany Ding
MTS · Statistics
Berkeley statistics PhD; conformal prediction — calibrated confidence for model outputs. The hire against one problem. Joined Aug 2026.
07
Sanjay Kairam
MTS · Evals
Stanford PhD; OpenAI, Reddit, Twitch. “Teaching machines to simulate humans.” Joined Jun 2026.
08
Mihika Kapoor
Head of Product
From Figma via HBS and Pear VC; hired Nov 2025, month five. Six Figma alumni followed her in 2026.
09
Chris Jensen
Founding Global Sales Director
Cresta, Splunk, SignalFx. Hired Feb 2026, the month of the Series A — the first of eight from Cresta and Medallia.
10
Madison Fox
Head of Deployments
Wharton MBA, Oxford; Bain & Company, PwC. Runs forward-deployed delivery from New York.

MTS = Member of Technical Staff.

4.3 — The shape of the team

A mid-career commercial crew, titled like a lab.

8.2
Years average experience — a mid-career crew, not new graduates and not veterans.
8.5%
Hold PhDs — 5 of 59.
75%
Senior level or above — by title.
80%
Hired in 2026 — 47 of 59.
49%
In the Bay Area — 29 of 59.

Engineers carry one title — Member of Technical Staff — and the rest of the company mirrors it: Member of Product Staff, Member of Design Staff, Member of Recruiting Staff. Underneath: 29 joined senior, 18 mid-career, eight veterans, four early-career. Eight MBAs to five PhDs, at a company founded to build foundation models.

Schools underneath: Stanford ×22 degree mentions, Harvard ×9, UC Berkeley ×8, Cornell ×3 — one dominant feeder school, unusual for the cohort. The company page's own facets over all 67 members agree on the shape: Stanford 15, Harvard 5, HBS 5.

4.4 — The build order · 2025 = the nucleus · 2026 = the organization

Twelve in 2025. Forty-seven in 2026.

Jun – Sep 2025
The nucleus. Park incorporates in June with a founding MTS from SAIL (Pondoc) and a Head of Data (Patel); Bernstein and Yallen in July; a Founding Engineer from Synthesia (White) in August.
Oct – Dec 2025
Commercial hires before the money. A Stanford HAI graduate into Deployments (Isabelle L.), the Head of Product from Figma (Kapoor), GTM leadership from McKinsey, the first recruiter. Twelve by year end.
Jan – Feb 2026
Series A — $100M from Index — and the sales floor opens. Founding Global Sales Director from Cresta (Jensen), Head of Deployments from Bain (Fox), two Founding GTM hires, three engineers (Figma, Glean, Apple).
Mar – May 2026
The Figma pull and the incumbent lift. Wymer (Cresta, Medallia) to run global sales; Stoy (Figma) and Catania (Notion) into product; the first Evals hire (Srinivasan); three deployers from Bain; a second recruiter.
Jun – Jul 2026
The research bench arrives, a year in. Kulkarni (Microsoft AI), Kairam (OpenAI), then Tony Lee (HELM) with two Figma engineers and five deployers. Nine hires in July; the Series B closes on the 31st.
Aug – Sep 2026
After the $200M. A Berkeley statistician (Ding), a Citadel engineer (Zhen), a DeepMind alumna (Skylar Wang); more sellers from Cresta and Medallia (Brown, Glazov, O'Meara); a third recruiter.

Park had planned to “go aggressive toward the end of 2026”; buyers “closed deals within 3 months,” so the floor opened in February. Through May, sellers and deployers out-hired builders nearly three to one.

4.5 — Where they came from

Two networks, one anchor apart.

The first ring is the lab: nine people carry a working Stanford affiliation. The second is Figma: a Head of Product in November 2025, six more alumni by July. Neither ring runs through a job board.

Figma · 7Stanford University · 6Cresta · 6Medallia · 5Microsoft · 5BCG · 4
The Stanford lab
Joon Sung ParkMichael BernsteinTony LeeChristopher PondocIsabelle L.Olivia WangSarah C.Sanjay KairamAndrew Wesel
9 of 59 · 15% · “my lab mates”
one anchor
Nov 2025
The Figma pull
Mihika KapoorJenning ChenJasmine StoyNatasha TenggoroKeeyen Y.Carina RamosRohit Chouhan
anchor Nov 2025 · six followed Feb – Jul 2026

Below the rings, two more groups that never touched a job board: Jensen · Wymer · Brown · Hittson and four more, lifted from Cresta and Medallia — eight sellers in all; and Zhen · Chen · Haarmann · Lauer, the quants, from Citadel, DRW, PIMCO, D. E. Shaw, ADIA and J.P. Morgan. Seven people carry a quant past in total.

4.6 — Who sells the model

The sales floor is real, large — and lifted from the incumbents.

01
Jensen · Wymer · Brown · Aronow
The incumbent's field force · Hittson · Glazov · O'Hara · O'Meara
Eight people from Cresta and Medallia, all hired in 2026. They sold contact-centre AI and survey research to the buyers Simile now calls.
02
Fox · Moonjely · Salzman · Quill
The consulting spine · Musuvathy · Zoe Y. · Boyd
Bain ×3, BCG ×4, PwC, Palantir; Wharton, five from HBS, a Schwarzman Scholar. Deployments is a consulting engagement with a model inside it.
03
Isabelle L. · Olivia Wang
The lab's own deployers · Stanford HAI · Symbolic Systems
Two Stanford graduates carrying the research to customers; the first Deployments hire came from HAI in Oct 2025, before the sales floor existed.
04
Smith · Vallurupalli · Lauer · Franca
GTM leadership · Instacart · McKinsey · Blackstone · Insight
A Head of Consumer from Instacart; McKinsey and Blackstone alumni; a former chief of staff to a CEO. Finance and consulting, not SaaS.
05
Boyd · O'Hara
The customer channel · CVS Health · customer and investor
A product lead into Deployments in July, a CX executive into GTM in April — from a customer that is also a Series B investor. The buyer joined the vendor.

Twenty-six of 59 sell or deploy. Against it: six people with a publication record, two of them founders — the bench that advances the model is a tenth of the company.

4.7 — The story and the roster

A research company on the tin. A sales organization on the roster.

26: 23
Sell-and-deploy hires vs build hires
Of 59 people.
8 > 5
MBAs vs PhDs
At a company founded to build foundation models.
0
Public open roles
Careers page behind a login — with three recruiters on staff.

— The roster · 59 people, by our read · Sep 2026

— The story · “founded by Stanford researchers to build foundation models” · who publishes

The company's description — foundation models, a Stanford lab — is true of a tenth of the roster. The other nine-tenths sell, deploy, design and staff a product the tenth invented. The enterprise organization is the company — by design: “I was not an enterprise seller,” Park says. He hired one.

§ 05 — Part five of five · What transfers

What the rest of us can steal.

You can't copy the advisors or the $2B. The direction of the recruiting, the anchor-then-cohort mechanic and the incumbent channel transfer.

5.1 — Sourcing patterns

Six patterns — none of them ran through a job board.

01 · Recruit upwardThe student hired his advisorsPark convinced Bernstein and Liang — “they were actually my doctoral advisors” — to join as co-founders. Nine of 59 then followed from the same lab. Inverted seniority as the founding act.
02 · Anchor, then cohortOne Head of Product, six Figma alumniKapoor arrived from Figma in Nov 2025; Chen, Stoy, Tenggoro, Keeyen Y., Ramos and Chouhan followed Feb – Jul 2026 across engineering, product and design. Figma out-feeds Stanford.
03 · The incumbent channelBuy the field force you're replacingCresta 6 and Medallia 5 — eight sellers, every one hired in 2026, from Founding Global Sales Director to Strategic Accounts. They already know the buyer.
04 · The consulting spineDeployments, staffed like a case teamBCG 4, Bain 3, McKinsey 2; eight MBAs, five from HBS. Forward-deployed delivery run by people trained to run research projects for clients. No public description of the company mentions it.
05 · Hire the instrumentsThe people who built the benchmarksHELM's builder (Lee), two “Evals” titles (Kairam, Srinivasan), a conformal-prediction statistician (Ding). When accuracy is the product, hire the measurers.
06 · Recruiters, no postingsThree recruiters, zero public rolesTommy F (Dec 2025), Guanzing (May), Hakimjavadi (Sep) — one every four months — while LinkedIn shows nothing open and the careers page needs a login. Network hiring, professionally staffed.

5.2 — The lessons

Five moves worth stealing.

01 · SourcingRecruit up the org chartFounders map their future reports; Park mapped his advisors. Ask who taught the founder, and whether they'd take the call — a professor as co-founder brings the lab behind them.
02 · Units of hireLand the anchor, let the cohort formOne senior product hire from a design-tool company drew six colleagues in eight months. Choose the anchor for the network they carry, not only for the role.
03 · The poolThe category you replace trained your sellersSimile's field force came from Medallia and Cresta — the companies that sold surveys and CX software to the same buyers. List the incumbents' sales teams before writing a sales req.
04 · SequencingSell in parallel, not afterPark planned to “go aggressive toward the end of 2026”; buyers closed “within 3 months,” so the floor opened in February — four months before the research bench. If the product is a paper, hire the contract-makers early.
05 · ScreeningUncorrelated superpowersTwo strengths that shouldn't coexist; success across several contexts; “married to impact, not a problem.” Screens for the hybrid roles this roster is full of.
The catchCash won't close the researchersPark can't match researchers' base pay, he says; six of 59 publish, and the retention claim waits for 2027. The research core is the hard part — hire it first.

— The takeaway

They turned a paper into a company.
Then they hired the sales floor
of the category it replaces.

Brief
Simile · Talent Brief
Prepared by
Base to Base · Recruiting
Written for
Founders turning research into a company

Methodology & limitations

Wide coverage — honestly counted.

— Sources

  • The LinkedIn company page (67 associated members), with full profile and education histories for 59 of ~60 identifiable staff, collected 12 September 2026.
  • Google Scholar profiles for the six members with publication records, fetched 14 September.
  • Three founder appearances with transcripts — Latent Space, 20VC, Sequoia Training Data — and a company film.
  • Funding disclosures and press.

— Limitations

  • Eight non-staff excluded from team statistics — three VC partners, two advisors, one angel, two name collisions.
  • Percy Liang is absent from the company page, and so from every roster ratio here — which undercounts the research bench by one. Karpathy's angel listing was only partially visible.
  • A 30-day public-signal scan (X · Reddit · Hacker News · GitHub) produced no usable result, and is recorded as such.
  • The 26 : 23 sell-and-deploy to build split is our classification of public titles, not the company's own. Teardowns like this are how our searches begin — this one's on us.