Handwired

Handwired is an open-source experiment by Cheng Lou on GitHub. It takes ordinary JavaScript programs built from if-else branches, loops and state, and rewrites them as matrix math using only addition, multiplication and ReLU, which makes them a neural network. The result runs on JavaScript or WebGPU. In the author's words it is "no training, exact, yet also fully differentiable": the weights are written by hand, so nothing is trained; the output matches the original program exactly; and because only those operations are used, the whole program supports gradients. What stands out is that the network's weights come straight from the program, with no training data at all. The author suggests unfreezing those weights to train them further, or attaching such a program to a general network like nanoGPT and training the two together. There are live demos, including a search over more than 100,000 JSON documents.

Who it's for:programmers and researchers curious about differentiable programming and how neural networks work. Readers who don't write code won't have a use for it, though the live demos are worth a look.

Before you start: to run it locally you need Bun; after git clone, run bun install and bun start. The live demos open in a browser. No account or API key is required. It's a personal experiment under the MIT license.

Based on the project's public documentation. Not tested by AI Inspiration World.