Every gathering is a simulation waiting to run.
Twenty five people, their work, their questions, the connections no one planned.
Give the room a model, and chance becomes design.
A world model learns enough about a situation to run it forward. Point one at a gathering and you can see the room happen before it does: who would meet, what would change, what a traditional format leaves on the table that a designed one captures. We are building it. The pieces are already on this site.
People cluster with who they already know. The valuable connection across the room is left to chance.
What a world model does for a room
The frontier labs use world models to run a scene forward: a car predicting the next second of road, a game generating the next room. Applied to a gathering, the same idea becomes a model of each attendee, built from their public and professional signal, that estimates likely interests and useful connections. Not a copy of a mind. A model that lets an organizer rehearse the room and design it on purpose instead of hoping.
The proof that this is real, not science fiction: a Stanford team put twenty five model agents in a small town and, unprompted, they planned and threw a party. A follow up built models of more than a thousand real people that reproduced those people's own survey answers about eighty five percent as well as the people themselves did two weeks later. Twenty five is where a gathering starts to need design. That is where we start.
An illustrative walk through
Take a real gathering already in the atlas: Learning Forward Annual Conference 2026, which scores 59 on the Gathering Effectiveness Score. The world model would read that agenda, then read a model of each attendee. Among the kinds of people the atlas already tracks are Steven Rogelberg (meeting science), Nancy Margulies (facilitation and visual thinking), and John S. Dryzek (deliberative democracy). In a traditional format, the people who most need each other meet only if luck seats them together. The model surfaces the bridge on purpose: the facilitator who can turn one person's research into a room's practice, the deliberation designer who can give another's expertise a decision to land in. It names the connection, the reason, and the exercise that would make it happen. This walk through is illustrative. It shows the shape of the output, not a claim about what any named person would actually do.
The pieces are already live
We are not describing vaporware. Every component of the world model is already built and running on this site. The frontier is fusing them into one model that runs a specific room forward.
What this model will and will not claim
- Opt in, always. A model of a real person is built only with that person's consent, from signal they approve. No one is simulated without opting in.
- We model likely patterns, we do not predict what you will do. Outputs are directional and probabilistic. A suggested connection is a hypothesis worth testing, not a forecast of anyone's behavior.
- Represent, never replace. A model stands in for a person's professional profile in one bounded context. It never impersonates them, speaks for them, or acts on their behalf.
- A model is not a mind. The honest research on this is clear that individual models get the average right and the surprising, human tails wrong. We say so, and we design for it.
- Scoped and deleted. A model is built for one gathering and retired after, not kept as a standing double of a person.
Start with the room you have
The world model is the horizon. The tools that feed it are free today.