The construction of a large language model (LLM) depends on many things: banks of GPUs, vast reams of training data, massive amounts of power, and matrix manipulation libraries like Numpy. For models with lower requirements though, it’s possible to do away with all of that, including the software dependencies. As someone who’d already built a full operating system as a C learning project, [Ethan Zhang] was no stranger to intimidating projects, and as an exercise in minimalism, he decided to build a generative pre-trained transformer (GPT) model in the kernel space of his operating system.
As with a number of other small demonstration LLMs, this was inspired by [Andrej Karpathy]’s MicroGPT, specifically by its lack of external dependencies. The first step was to strip away every unnecessary element from MooseOS, the operating system [Ethan] had previously written, including the GUI, most drivers, and the filesystem. All that’s left is the kernel, and KernelGPT runs on this. To get around the lack of a filesystem, the training data was converted into a header to keep it in memory — at only 32,000 words, this was no problem. Like the original MicroGPT, this is trained on a list of names, and predicts new names. Due to some hardware issues, [Ethan] hasn’t yet been able to test this on a physical computer, but it does work in QEMU.
It’s quite impressive to see such a complex piece of software written solely in C, running directly on hardware; for a project which takes the same starting point and goes in the opposite direction, check out this browser-based implementation of MicroGPT. For more on the math behind GPTs, check out this visualization.
Continue reading “Building A Dependency-Free GPT On A Custom OS”
