GPT-2 implemented in pure CMake runs at 47 tok/h

AlpinDale · x · 2026-08-24

A developer implemented the GPT-2 model entirely in pure CMake, demonstrating that the build system is Turing-complete. The model runs using Q16.16 integer arithmetic at a very slow speed of 47 tokens per hour.

Original post →

More from Fun

Fun channel →