recall from noise
A Hopfield network is content-addressable memory. Store a pattern, feed it a corrupted version, and it relaxes back toward what it remembers.
STORE
The character bitmap becomes a vector of ±1 cells. Hebbian weights W[i][j]=A[i]·A[j] etch it into the energy landscape.
CORRUPT
Random cells flip — the pattern is buried in noise, sitting up the slope of its energy basin.
RELAX
Over 8 passes each cell re-checks its neighbours, with a bias λ=0.35 toward the stored state, rolling downhill.
RECALL
The network settles at the basin floor — the original character, reconstructed. The memory is recovered.
the screen treatment
Nothing is a flat pixel grid. Every frame renders at 2× and runs a five-stage CRT post-process, sampled bilinearly so the curved text stays sharp.
BARREL
Edges curve outward like a real tube.
VIGNETTE
Corners darkened for depth.
ABERRATION
Red/blue fringing near the edges.
SCANLINES
Stronger at the edges, soft in centre.
GLOW
Phosphor bloom rolling off to the corners.
every character sings
Each piece carries its own generative score, built live in the Web Audio API — no samples, just oscillators, reverb and a saturator. The composition is keyed to the character: E sings in E major, + builds an additive chord, the space resolves into silence.
Corruption stirs a low wash; each recall pass bleeps a note; convergence rings a chime in the home key.
record your specimen
Pick a character, watch it corrupt and recall, and capture the loop as a video — full CRT treatment baked in. Your own trailer to keep or share.