See how a society will move
before it moves.
Language models predict the next word. World models predict the next frame of the physical world. A social world model predicts the next state of a society — people, their relationships and the media between them. Drop a product launch, a policy or a rumor into a synthetic society of real personas, and watch who hears first, who they tell, and where opinion splits, day by day.
The third kind of predictive model: society
A survey captures what people think at one moment. A social world model captures time — how those thoughts change as they travel through relationships. Who knows first, whom they listen to, and where the split happens.
Language model
- Predicts
- The next word
- Input
- Text
- Used for
- Writing, summarizing, conversation
World model
- Predicts
- The next frame of the physical world
- Input
- Video, sensors, actions
- Used for
- Robotics, autonomy, physical AI
Social world model
- Predicts
- The next state of a society (opinion, behavior)
- Input
- Population, network, media, events
- Used for
- Rehearsing launches, policy and crises
Seven real societies, one engine
Each society in the demo is built from these numbers. The people are real synthetic personas grounded in each country's census statistics; the economic mood comes from the latest World Bank indicators.
Loading data…Economy World Bank, latest year available
People
Age
Region
Top occupations
Drop an event into 800 real personas
Each dot is one person — tap it to see who they are. People cluster by region, and lines are acquaintances. The same society runs twice so you can compare without intervention and with intervention side by side.
This world now
30-day trajectory solid: with · dashed: without
Stance by age
How much people only know others of the same region and age. Higher means news gets trapped in islands.
How strongly ads, articles and broadcasts reach people. 1 is the baseline.
How much people doubt the message. Higher means less movement and more backlash.
Draws a new set of relationships. See how much the outcome wobbles.
Reading this result
This demo is a rule-based miniature that runs in your browser. The people (age, sex, region, occupation) come from real persona datasets, but their judgement is imitated with simple rules. The numbers are for getting a feel for comparing scenarios — they are not forecasts of real public opinion. In the full engine, a language model takes on each persona and decides.
One event, seven societies
The same event and the same intervention, dropped into each society with the settings above. Where does it catch on, and where does the intervention matter most?
Day-30 result. Top bar: without intervention · bottom bar: with intervention. Select a row to open that society in the demo.
A society, built in four layers
Each layer brings the simulation closer to the real thing. The demo has a simple version of all four; the full engine fills each one with data and language models.
Population
People with an age, sex, region, occupation and way of life — personas matched to each country's statistics.
Demo 800 personas per country sampled from NVIDIA Nemotron-Personas (CC BY 4.0)
Network
Family, coworkers, neighbours, online friends. Similar people connect; a few hubs reach far.
Demo Region and age homophily + 3% hubs
Media
Who social media, TV and news reach, and how hard. Younger people lean social, older people lean broadcast.
Demo Per-channel exposure · calibrated by internet use, growth and inflation
Judgement
People who hear the news weigh it against their own leanings and their neighbours' views, decide, and pass it on.
Engine Each persona decides through a language model, with deliberation rounds
Rehearse before the decision you can't undo
Market response before launch
Which ages and regions catch fire first, and where it stalls. When a creator program travels furthest. How the same launch lands in different countries.
Should the creator program start in week one or week two?
Policy and program acceptance
How opinion splits after an announcement, and whom to brief first so opposition doesn't harden.
If we hold town halls for seniors first, how much does opposition drop?
Crisis response rehearsal
Which relationships a rumor travels along, and how much distrust sets in for each day a response is late.
What do we lose if our statement comes three days late?
Organizational change
Drop a new policy into your company's own network and find who to persuade first.
Which of our 20 team leads should we talk to first?
From surveys to societies
The social world model is being built on top of an AI polling engine that is already running in Korea.
1. AI polling
30–300 personas matched to your target audience answer a survey; results are broken down by age, sex and region.
Try it (Korean) →2. Deliberation rounds
Everyone sees the first-round results and is asked again — measuring how much opinion moves once people see what others think.
Turn it on in AI polling →3. Social World
Personas linked into networks, with media and events, comparing multi-day spread and intervention effects across countries.
Join the PoC →Want to run your own event?
We'll define a launch, policy or crisis scenario with you — in the countries you care about — and run it on the full persona engine as a proof of concept.