Computer week 2026 · Monday night

World first

frame 0 · micro-step 0

About this control

Click any screen to empty a disk of gas in all five at once. Show the camera's 32 × 32 blocks Hide the names Training films

Archive progress

Measured results

The registered run

Twelve predictions were committed before any of this code existed. The run used 200 training films, 20 of those films replayed, 20 new films, the word WORLD, and a slit.

What the numbers say

One frame at a time the models are hard to tell apart: every array of counts is the picture of some state, and the mosaic's counts match the world's distribution to a total variation of . The difference is in how frames follow one another.

The mosaic stays with the world for about five frames and then leaves it, on films it was built from as much as on new ones. Recall replays a training film exactly and ignores every edit. The world from a picture keeps the law and loses the particles' velocities, so it stays close without being the same world, and it cannot run a picture back to the word. Only the world run from its own state returns the letters, bit for bit.

The rule learned from particles is exact on every new film and on the word it never saw, and walks straight through walls it never met. Learning made a world where the model's state was the world's state.

What failed

Limits

One world (the HPP lattice gas), one camera (particles per 4 × 4 block), one training distribution.
Recall and mosaic are the optimal models of two architectures on the training films, not trained networks.
The learned rule sees particles, which no camera shows.
This page trains on the first 50, 100 or 200 of the registered training films, so with 200 its models are the registered ones.