Autoregressive Language Model on the 6502 Processor
Summary
This post documents a retro-computing AI project that fits and runs an autoregressive language model on a 6502-based BBC Micro, using a tight memory budget (9KB inference code, 13KB weights within 32KB RAM). It covers BitNet quantization, a 4-parameter-per-byte packing scheme, and a C-based inference kernel, with emulation and browser-based testing. It highlights hardware-aware ML design and demonstrates running a small LM on vintage hardware.