Fabrice Bellard: Fast, Standalone CPU Inference for Transformer Language Models

also see https://bellard.org/nncp/ Lossless Data Compression with Neural Networks; which appears to be the immediate prompt.