Who chooses what a world model builds

Two kinds of model

One group of systems models the world. World Labs made Marble generally available in November 2025; it builds editable 3D worlds from text, images, video or rough layouts. Google DeepMind’s Genie 3, shown in August 2025, generates worlds to explore in real time, and Project Genie opened it first to Google AI Ultra subscribers in the US in January 2026. Runway released its general world model, GWM-1, in December 2025, and Yann LeCun’s new lab, AMI Labs, is building models that predict how the physical world changes.

The other group models the person. Assistants now remember past conversations, read a person’s mail, photos and writing to answer in context, and are being built to feel tailored to each user.

The missing choice

Neither group decides what to build for one person. Marble’s own announcement says it “must invent all the details of the world that are not present in the input”. Someone has to supply those details. A person’s taste is not only what they already like. It includes how far from their own usual they are willing to go, and which surprises they keep.

LeCun’s 2022 blueprint for machine intelligence already separates the two jobs. A world model predicts what would happen, and a separate cost module scores how good each outcome is. Our work belongs on the second side, for one person: generate many ideas, place each by how far it sits from the usual, and learn which ones this person keeps.

Where our work fits

Our ideas become drawings and a Rhino model in the studio. Our tests show that a person’s lean toward the unusual is only a small part of their taste (A selector that learns one person), so the chooser has to learn more than distance. What we would add to a world model is the choice of what to build, and a measure of whether that person keeps it.