Building A Dependency-Free GPT On A Custom OS

Ethan Zhang built KernelGPT, a dependency-free GPT model running entirely in the kernel space of MooseOS, a custom OS he previously wrote in C. Inspired by Andrej Karpathy's MicroGPT, the project strips MooseOS down to just the kernel, embeds training data (32,000 words of names) as a C header file to avoid needing a filesystem, and runs the transformer model with zero external libraries. Currently tested only in QEMU due to hardware issues, it demonstrates that a functional LLM can be built and run in bare-metal kernel space using pure C.