The living observatory

Drag to explore. Tap something to give it a push.

A living neural engine

The recurrent network advances every tick. Trainable 3D fields describe each object. A neural material model colors the surfaces.

0neural ticks

—simulation weights

—GPU ms / frame

01 / LEARN HOW TO MOVE

Teach the world a new force law

Start with weak springs. Fit five neural weights to recorded trajectories. Six trajectories train the model; two are held out.

Training runs on your device. The world keeps moving.

02 / LEARN WHAT EXISTS

From spheres to objects

The same local feature fields before and after learning 3D shape and color observations.

A checkpoint comparison. New text-created objects are compiled into weights and calibrated by gradient descent.

EXPERIMENT / LEARN FROM FRAMES

One loss, motion and appearance

A differentiable field renderer fits spring weights and object color together from synthetic RGB frames. A third camera and unseen times test the result.

See the fit ↗

Known shape, camera and action inputs. No state labels in the training loss.

03 / INSPECT THE BOUNDARY

Language proposes a world program. Existing object identities and neural state survive edits. For new objects, pretrained image and 3D models supply observations; gradient descent fits a compact neural field. Motion continues through the recurrent model.

Text creation uses the connected research GPU. Once loaded, the world and motion training run on your device. Camera rays, ray traversal, field interpolation and light-feature construction use conventional GPU code. This is a hybrid neural engine, not a video diffusion model or an unrestricted learned world simulator.

Results & limitations ↗Source ↓

Latest world program

No text edit yet.