Handwired

chenglou/handwiredReviewed by AI

What it does

Handwired turns plain JavaScript logic (if-else, loops, state) into additions, multiplications and ReLU, then folds it all into one big matrix multiplication that runs in JavaScript or on the graphics card through WebGPU. No training is involved: the numbers are written by hand, and the result stays differentiable, so the author explores training it or bolting it onto a real neural net. It is for programmers curious about the idea. You can try the demos on the author's web page, or clone the repository and run it with Bun.

How easy it is to install

E · developers only · Windows

There is nothing to install for an end user: you either open the author's demo page in a browser, or clone the repository and run it with Bun and read the code in docs/, src/ and template/.

Coming soon: install in 1 click with the RepoPicks app.

Numbers on GitHub

Stars: 2,087on Oct 31, 2026
Downloads: —all releases
Growth: +58stars in 4 days

Stars over time

Facts

Runs on: Web, Windows, macOS, Linux
You need to install: Node.js, Other tools
To build from source: Node.js
Disk space: 100 MB–500 MB · estimate
Accounts or keys: None
Cost: Free
License: MIT (Permissive)
Last update: Oct 17, 2026

Risk signals (3)

Infoexperimental project
The README frames it as an exploration: the matmul runs "on JS or WebGPU, for fun", ideas are posed as questions ("what does that even do? Let's find out!"), and lab/ is described as "a bunch of old experiments you can check or ignore".
AI review
Infonot published as a package
package.json sets "private": true and there are no GitHub releases, so you cannot install it from npm; you have to clone the repository.
AI review
Infoneeds bun not node
Every script in package.json runs through bun (for example "start": "bun scripts/serve-demo-site.ts"), and the README says `git clone && bun install`, so the Bun runtime is required rather than plain Node.js.
AI review

Evidence is shown in its original language.

What Radar could not check

Whether Bun and the demos run without trouble on Windows.
Whether the WebGPU demos work on your own graphics card and browser.
How well the idea holds up on programs larger than the included demos.
How much disk space a clone plus its dependencies takes.