It isn’t. The game runs on a CHIP-8 emulator I wrote in C. The opponent is a neural network I trained with my own C library.
Here’s how it worksForward and backward pass, mini-batch SGD, activation and loss functions, no dependencies. The first version took two days and worked on flat 1D arrays. As a sanity check it reaches 97.6% on MNIST.
A game agent is never told the right move. With REINFORCE the network outputs a softmax over actions, and the actions that led to a good return become more likely.
4 KB of memory, 16 registers and a 64×32 display, written in C, with a compiler I created that compiles assembly to binary CHIP-8 opcodes and a browser editor. Pong runs on it as a ROM.
The emulator steps the game, MLKit reads it and presses a key, and the outcome of the rally flows back as the training signal. One C program driving another.
Both C codebases compile to WASM. The emulator lives in Worker A, inference in Worker B, pretrained weights load once, and the page thread only paints.
Programming lessons in Polish, with an in-browser editor and automatic solution checking. I built the architecture, backend and frontend, including sandboxed C++, Rust and Go through Docker-to-WASM builds. Silver Wolf winner, mentored by Allegro.
A game-making environment for CHIP-8 in the browser. You write assembly; a compiler I wrote turns it into CHIP-8 opcodes, and a C emulator core running as WASM executes it with a live debug panel. It also powers the Pong game above.
A neural-network library from scratch in C: backpropagation, mini-batch SGD, activations, losses and REINFORCE. 97.6% on MNIST. It is the brain of the Pong opponent above.