
Only the Gradient
Draw on a wall of 256 neurons; a server prints your drawing back from the gradient.
I build AI agents and the systems behind them, keep them running, and evaluate their performance.
Arrow keys spin the globe, plus and minus zoom, and zero goes back to the start. Tab moves on to the project satellites.
Off by default. When it is on, each satellite chimes in its own note as it passes nearest you, and a ring of light shows each chime.
RUNS ON YOUR DEVICE
Each one runs on your device and says plainly what is computed live and what was recorded beforehand.

Draw on a wall of 256 neurons; a server prints your drawing back from the gradient.

A tiny network keeps five notes in two neurons. You play it as five strings of light.

Pour sand into two vessels: one follows the real rule, one a network that learned it.

Drag the pressure: painted specimens curdle into neon as a reward model’s score climbs.

A clothing classifier’s neurons as glowing pictures. Take their bits away two ways.

A mammoth made of light, pressed flat four ways. Each way tears it somewhere else.

Ten AIs rewrite a page of Moby-Dick, each taught only by the last. Rare words go first.

One graphics card’s memory, lit up: the model takes a block, every conversation the rest.

An AI email assistant in a lit box. Hidden lines fool it; the box stops every send.
Night Earth: NASA Earth Observatory, Black Marble 2016, from Suomi NPP VIIRS data (public domain). Drawn with three.js. Type: Inter, JetBrains Mono and Instrument Serif, self-hosted. No analytics, no trackers.