Every demo below runs the real Handwired runtime in your browser — no server, no GPU required. Upstream demos link to Cheng Lou's site.
## Which demos run on this site?A search formula learns its weights from 1,000 shoppers while its business rules stay exact; a generic net on the same clicks breaks them.
Drag cards to the sizes you want; the sizing function learns its constants and nothing overflows.
Fit gravity, drag and bounce to a recorded path, then replay it as code.
A theme learns from a designer's picks while WCAG AA contrast stays exact; a generic net on the same picks fails it.
Upstream's 100,000-document search, embedded with credit: GPU timings next to plain JavaScript.
A 7-step lesson. A tiny JS rule becomes an exact neural net; try to break it, then watch it learn new constants in its own code, next to a regular net.
The README's a > 4 ? 10 : 12 as two ReLUs and a select. Editable a, branches and live plot.
Pick a relation, toggle the integer grid, fuzz 5,000 random pairs against plain JS.
Visualize the operations and layers behind a select, a comparison and an exact clamp.
Start from the exact program, then fit its threshold, slope and branch values by gradient descent.
Upstream's template/ app: a spring-animated switch run as one numeric program.
A small desktop with three windows and a dock — focus, stacking, dragging, resizing, springs, a photo gallery — run by one program: 234 source cells, 4,646 live operations in 130 layers, 9,909 weights, zero training.
A photo layout drawn by a small network inside a program (6,320 operations in 8 layers, 244 trained source cells), trained with Adam using the program's own gradients while you edit the sketch.
Regex matching, JSON parsing and highlighting written as numeric programs, then emitted as WebGPU shaders that search every document at each keystroke.