Handwired compiles JavaScript apps into differentiable neural nets you can run backwards
Vjeux · x · 2026-10-18
@chenglou released a toy project called Handwired: it turns regular JavaScript code into a neural net — no training, exact outputs matching the JS program, and fully differentiable, all running in a browser without a GPU.
User anshuc built a handwriting-input calculator on top of it: the app looks like a normal web app but is actually a single neural net running every frame, simultaneously rendering the UI and recognizing handwriting. The recipe: write the calculator in JS, compile it to an NN with Handwired, then splice a handwriting recognition model directly into the weights.
The truly cursed part: since the whole app is an NN, it can run in reverse — start from an answer, and gradient descent flows backwards through both handwriting and app logic to find inputs that produce that answer.
More from Fun
- Waymo's pumpkin Halloween-wrapped robotaxi spotted on LA streets — omooretweets · 2026-10-18
- "Mass AI psychosis": arohan mocks employers that don't hand out spirit animals — _arohan_ · 2026-10-18
- GPU legend Satnam Singh unearths his 2009 Microsoft Research paper during garage clear-out — satnam6502 · 2026-10-18
- Indian restaurant chain's menu goes AI-slop: bizarre triangular dosa images baffle customers — andrewgwils · 2026-10-18
- Miles Brundage jokes about a 'psyop' claiming the Bay Area has no all-night restaurants — Miles_Brundage · 2026-10-18
- Garage clear-out of classic software engineering textbooks sparks 'aged poorly' commentary — satnam6502 · 2026-10-18