Your if-else is secretly a neural net (and it never needs training)

What it does

Handwired takes regular TypeScript/JavaScript — branches, loops, state — and rewrites it as matrix multiplication plus ReLU over floats. The trick is that comparisons become arithmetic: a > 4 turns into a difference of two clamped terms that yields exactly 1 or 0, and that indicator then blends both branches of a conditional instead of picking one. Everything collapses into one giant matmul that runs on plain JS or WebGPU.

The interesting bit

Because the entire program is built from just +, *, and ReLU, it’s fully differentiable — yet the weights are hand-written, not trained. The result is a “neural net” that computes your program exactly. The README then cheerfully suggests the obvious mischief: unfreeze the weights and train them anyway, or bolt a frozen program onto a real model like nanoGPT and train end-to-end.

Key highlights

• Exact program semantics with zero training — weights are derived, not learned.
• Fully differentiable, so you can attach it to a trainable network.
• Runs on JS or WebGPU; the minimal primitive set is friendly to fusion, sparsity, and reordering.
• A JSON search demo reportedly scans 100k documents in milliseconds, thanks to GPU-style parallelism from single-threaded-looking code.
• Live demos are hosted on the project page.

Caveats

• The README is a proof-of-concept in spirit: the “funny use-cases” (training the weights, nanoGPT integration) are posed as open questions, not shipped features.
• The comparison-to-arithmetic encoding shown works for integers; how it generalizes to all real-world JS isn’t spelled out.
• A lab/ directory of old experiments hints at plenty of churn behind the current design.

Verdict

A delightful piece of program-compiler art from chenglou — worth an hour of poking if you care about the boundary between code and models. Not a library you’d ship to production tomorrow.