Is this the thing you meant?
WaitingPick a subject.
Type anything. Tens of thousands of points redraw a picture of it as a relief you can turn, and a judge scores how close they came using only what the points draw.
Pick a subject.
A separate straight-on camera with no perspective renders only the points at 256 px, whatever angle you turn the stage to, so the relief flattens out. That render is compared with the source on black using SSIM on red, green and blue, where 1 means identical. Patches that are empty in both are skipped. The grey rows are the same judge applied to simple stand-ins built from the source, so the number has something to stand against.
Sends only a render of the finished points to Claude, without the subject's name, and asks what it shows.
Points are drawn at even steps along a space-filling curve through the picture's weight, so they spread out without clumping. Each point's size answers how crowded its neighbourhood is.
Parts brighter than their surroundings sit closer to you. From the front it reads as a picture; turned, it reads as terrain.
Arrival is ranked by edge strength and darkness, like a draughtsman. Unmaking lets the light leave first.
The likeness number comes from a render of the points, never from whether they reached their targets.
Search sends your words to Wikipedia and uses its article ranking and lead images, with Wikimedia Commons as a fallback. Photos you add are read on this device; one is only sent anywhere if you press “Ask Claude”, and then only as the points render. The page loads fonts from Google Fonts and its 3D library from cdnjs. Algorithmic philosophy: Sworn Light.