chenglou/handwired is a project that converts ordinary JavaScript code into a neural network without any training. It expresses programs with conditionals, loops and state as neural network operations, and runs them in JavaScript or on WebGPU. It is written in TypeScript and uses the MIT license. The repository shows 1692 stars.
The README explains that the conversion uses matrix multiplication and the rectified gated linear (reglu) operation. Instead of running only one side of a conditional, it computes both and picks the result with weights that are 0 or 1. The example condition a>4 is expressed as Math.max(0, a-4) - Math.max(0, a-5). With these basic operations alone, the whole program becomes differentiable, and the weights can either be set directly or treated as training targets. It is also possible to connect a fixed program to an ordinary neural network and train them together. The README includes a demo that processes 100,000 documents in milliseconds via JavaScript search.
However, from the README alone it is unclear whether the project is a demo or a research effort. Differentiability and correctness are only explained through examples, and no formal verification figures or procedures are given. So the demo shows an interesting possibility, but the reliability and applicable scope of the approach cannot be judged from the documentation alone.