Worlds

World models can be built anywhere there is a changing world to learn. These are the first places we have worked, not the last.

Worlds for work

Clinical and regulatory development

Life Sciences

Follow a drug program through trials, submissions, approvals, label changes and withdrawal risk.

Current work: Project Confirm forecast FDA oncology outcomes across 3,275 held-out windows from drug programs the model had never seen.

Repositories and AI work

Software

Learn how work moves through a repository or an agent team, then predict what needs attention next.

Current work: Project OSS learnt from 7.04 million events and transferred zero-shot to a ninth repository. World Router uses the same idea to allocate AI work.

Anomaly and threat discovery

Cybersecurity

Learn what normally happens across a system so the few events that matter stand out from the millions that do not.

Current work: one model evaluated 1.1 million events and surfaced the 250 most unusual at 94% accuracy.

Worlds to play

Playable worlds

Games

Let a learnt model judge how each move changes the state of a game, rather than scripting every consequence in advance.

Current work: Mandate of Heaven turns six years of real AI history into a strategy game judged by a world model trained on 10,066 events.

Stories and invented worlds

Fiction

Model the state of a story so authors and agents can ask what is building, what comes next and what a change sets in motion.

Current work: Torus models a 100-scene manuscript. Star Wars and Middle-earth provide fresh fictional worlds for hidden-future tests.

The worlds are different. The underlying job is the same: build the record, learn how its state changes, and ask the model what happens next.

See the scored results