PYTHON · NUMPY

A playable world

Build a pixel-space, action-conditioned world model and play it — the exact principle behind Genie, Oasis, and our on-device world model.

01Challenge

Build a world model that predicts the next FRAME from the current frame and your steering action, then play it as an interactive rollout that stays coherent.

02Model

The frontier of world models is visual and playable. Genie generates an interactive world from a prompt; Oasis renders a Minecraft you can play, one AI-painted frame at a time; DIAMOND trains agents inside a diffusion world model. All of them do one thing: read the current frame plus your action and paint the next frame — then feed it back, so you can play. The Institute's own nano-world-model does exactly this, live and on-device, at /research/world-model. Here you build the smallest honest version: a blob you steer on a grid, a model that reads the frame, finds what matters, moves it by your action, and re-renders. Feed each painted frame back and you are playing a learned world with no physics engine underneath.

03Guided practice

1. A visual world. A blob on a grid you steer with an action; the real world renders the next frame. 2. Frame in, frame out. The model locates the blob, moves that estimate by your action, and re-renders — action-conditioned next-frame prediction. 3. Play it. Feed each predicted frame back in and steer; it stays coherent. That feedback loop is Genie, Oasis, DIAMOND, and our on-device nano-world-model, in miniature.

04Feedback

PASS WHEN PASS when your played rollout of predicted frames stays coherent against the real world.

If your run fails, check:

A cell errored — run them in order.
The played frames should track the real world's frames.

05Retrieve & space

What does a pixel-space world model map from and to? (from the current frame and an action, to the next frame).
What makes it 'playable'? (each predicted frame is fed back in, so your actions steer an imagined world).
Which real systems run this loop? (Genie, Oasis, DIAMOND — and the Institute's on-device nano-world-model).

06Mastery & project

You can build and play an action-conditioned world model, the principle behind playable world models.

Try the live version: /research/world-model runs this same loop as a learned neural simulator, on your device.

GO FURTHER · THE LIVE PLAYGROUND

Play the on-device world model ↗ — the same action-conditioned loop, running live as a learned neural simulator on your device

← Plan by imagining Certify, don't just score →

PYTHON · NUMPY

A playable world

Build an action-conditioned next-frame model and play it as a rollout.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you