Repolode
Submit
kerlenton

kerlenton/bootgpt

A tiny char-level language model that runs inside a 512-byte x86 boot sector and dreams up fake Linux boot logs in real mode.

View on GitHub
Stars
99
Forks
4
Open issues
0
Last commit
7 days ago

What is it?

bootgpt is a neural network that runs at boot time, before any operating system loads. The 512-byte MBR contains hand-written x86 assembly that reads an 11.5 KB int8-quantized model from the following 23 disk sectors and performs inference in 16-bit real mode on a 386-class CPU, with no OS or C runtime involved.

The model is a small character-level network with char embeddings, one hidden layer and a softmax, trained on about 3 MB of synthetic dmesg-style logs generated by a Python script. Training uses power-of-two quantization scales so the assembly can do the forward pass with integer multiply and divide. A Python script simulates the same integer arithmetic, and its output matches the boot sector's output character for character.

The project is mainly a demoscene and sizecoding experiment. It is meant to be run under QEMU, with a script that records a demo GIF from headless QEMU, and the disk image can also be written to a USB stick for a BIOS machine.

Key features

  • 01Boot sector code fits in 512 bytes and loads model weights from the following disk sectors.
  • 02Inference runs in 16-bit real mode using integer-only arithmetic.
  • 03Python training script quantizes the model to int8 and simulates the assembly's math for comparison.
  • 04Generates synthetic kernel-style boot logs as training data.
  • 05Runs under QEMU, and the disk image can be written to a USB stick.

Use cases

  • Studying how neural network inference can be written in raw x86 assembly.
  • Learning sizecoding and boot sector programming techniques.
  • Experimenting with int8 quantization and fixed-point inference on very constrained hardware.

Quick start

  1. 1

    Install nasm and qemu with Homebrew

    brew install nasm qemu
  2. 2

    Create a Python virtual environment and install numpy

    python3 -m venv .venv && .venv/bin/pip install numpy
  3. 3

    Build the boot image

    ./build.sh
  4. 4

    Run the image in QEMU

    ./run.sh

Things to know

  • The README's install commands are for macOS (Homebrew), so other platforms need their own setup.
  • The model is a tiny novelty that generates fake boot logs, not a general-purpose language model.
  • Booting real hardware requires a BIOS machine and writing disk.img to a USB stick, which the README only mentions briefly.

Topics

#x86#assembly#bootsector#language-model#quantization#qemu#demoscene#sizecoding#machine-learning#neural-networks

FAQ

What is bootgpt?
A tiny char-level language model that runs inside a 512-byte x86 boot sector and dreams up fake Linux boot logs in real mode. bootgpt is a neural network that runs at boot time, before any operating system loads. The 512-byte MBR contains hand-written x86 assembly that reads an 11.5 KB int8-quantized model from the following 23 disk sectors and performs inference in 16-bit real mode on a 386-class CPU, with no OS or C runtime involved.
Is bootgpt free to use? What license does it have?
bootgpt is open source under the MIT license. Check the license terms before using it in commercial products.
Is bootgpt still maintained?
Actively maintained. The last commit to bootgpt was on Oct 31, 2026.
What can I use bootgpt for?
Studying how neural network inference can be written in raw x86 assembly. Learning sizecoding and boot sector programming techniques. Experimenting with int8 quantization and fixed-point inference on very constrained hardware.
How do I install bootgpt?
Install nasm and qemu with Homebrew: brew install nasm qemu Create a Python virtual environment and install numpy: python3 -m venv .venv && .venv/bin/pip install numpy Build the boot image: ./build.sh Run the image in QEMU: ./run.sh
Where can I get bootgpt?
The source code and installation instructions are on GitHub: https://github.com/kerlenton/bootgpt

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects