Cheng Lou's Handwired compiles JavaScript into an exact neural net

On 15 October 2026 at 02:24 UTC, Cheng Lou posted what he called his latest toy project: Handwired, to “turn your regular JavaScript code into a neural net.” The post lists no training, exact outputs versus the JavaScript program, and fully differentiable, “runnable in a browser, without GPU.” His X profile says he worked on React, Messenger, ReasonML, and ReScript, and lists Midjourney, Pretext, and Freerange as current. The project page repeats that deck. The page itself does not show a publication time. The post is the date.

What the compiler is described as doing

The page’s example is a branch. For a > 4 ? 10 : 12 it writes a gate:

relu(a - 4) - relu(a - 5)

with ReLU defined as Math.max(0, x), then above4 * 10 + (1 - above4) * 12. Both branches are computed, and above4 picks one. The README walks the same example and says that for an integer a, that difference is 1 when a > 4 and 0 otherwise. It says a program made of if-else, loops, and state is transformed into matrix multiplication plus “reglu” over floats, the additions and multiplications and ReLUs are collected into one matmul, and the result can run in JavaScript or on WebGPU. The page and its diagram say ReLU. The README’s word is “reglu.” This account did not run the compiler, so it does not resolve that spelling.

The README says the weights are written by hand, so they do not need training, and that the program stays differentiable. That is the author’s description, not a measured error rate. The page and the README have no benchmark table.

Three demos, and the numbers on the page

The page lists three demos.

Mini OS is “a small desktop written in JS, transformed with Handwired,” covering lists, animation, stacking, scrolling, drag, input, and pictures, rendering to the ordinary DOM. The card’s figures are 9,909 weights, 130 layers, and training 0.

Learned gallery is one sentence: “Unfreeze a compiled program’s weights & train them in real time.” No weight count is given for that demo.

Search 100,000 JSON documents says the regex matching, the JSON parsing, and the highlighting are written in JavaScript and then turned into WebGPU kernels. The index card states no time. The README describes that demo as running in milliseconds over 100,000 documents. That timing is the README’s phrase. It is not a figure on the card.

The README also floats a further use, as a question rather than a result: attaching a frozen program to a generic neural net and training end to end, “now a nanoGPT can do both Shakespeare and your particular UI?” Nothing on the pages opened here shows that experiment as done.

License, and how the README says to run it

The LICENSE on the main branch of chenglou/handwired is the MIT License, copyright 2026 Cheng Lou. The README says to clone, run bun install, then bun start. Those are install instructions. This account did not run them.

What the toy does and does not claim

Handwired, in the author’s words, compiles ordinary JavaScript into a net that matches the program without training and stays differentiable, and the mini OS on the page is 9,909 weights and 130 layers with training counted as zero. One demo targets WebGPU. The announcement says the project also runs in a browser without a GPU. He calls it a toy. There is no speed table, no accuracy table, and no claim that this is a way to ship a product.