Doomformer compiles integer neural networks into ZDoom maps, using voodoo dolls for arithmetic and sector light levels for data storage. The system runs neural networks up to 65,000 times slower than NumPy. It can classify MNIST digits, generate text, and enable gameplay control.
A handwritten digit takes 163 seconds to process in real-time DOOM at default settings, or 8 seconds with optimization level 3. The project includes tools for drawing digits, checking computations, and playing DOOM with AI-generated maps. Installation requires Python 3.11+ and Linux or macOS, with a Colab notebook available for browser-based use without setup.
Source: Hacker News · Summarized by HeadlinesBriefing