The People Model — a new AI category, defined at Vurvey Labs.
The category triangle
Language Models
- Trained on text
- Answers with tokens
- GPT, Claude, Gemini
World Models
- Trained on physics
- Predicts what happens next
- Genie, Sora-class simulators
People Models
- Grounded in real human voices
- Reflects who people actually are
- People Model · LPM (Large Persona Model)
Two model families are already named. The third is the one that matters for every organization that serves people. The People Model — the Large Persona Model — grounded in real human voices, not scraped text or simulated behavior.
The compositional grammar
How a People Model composes.
Vurveys
It starts with a Vurvey — an opt-in video answer from one real person. Every People Model in the world starts with somebody choosing to answer.
From perspective to product
Workflows
Workflows are how multiple Agents compose new knowledge together — the way a good research team does.
Capabilities
Capabilities are the toolkit of things Vurvey does. Ready to be used solo, chained into Workflows, or delivered through Agents.
Agents
Agents are Personas configured to act — as research subjects when you want to interview them, as working assistants when you want them on your team.
Worlds
Worlds are the horizon. Populations composed into whole markets and cultures rendered through the people who live in them.
Look before you launch.
The dream of a "what-if machine" is not new. Sociologists and simulation researchers have chased it for fifty years. Early attempts — algorithmic agents, scripted characters — were rigid, and the academic literature delivered a stark verdict: "the models have been highly stylized and have had minimal impact."
What changed is what the models are made of. When digital twins are built from rich, individual data — long-form interviews with real people — they replicate survey responses with 85% of the accuracy that people replicate themselves. Twins built from stereotypes and demographics get to 70%. The gap is what we live in.
And when researchers stress-tested Vurvey-style agents against five published scientific studies, the agents replicated four. The fifth — the one the agents got wrong — turned out to be bad science that real people didn't replicate either. The simulation didn't just imitate people. It caught the study that shouldn't have been trusted.