BootGPT is a specialized project that implements a neural network capable of running during the initial boot sequence of a PC. Unlike modern software that relies on an operating system or a C runtime, BootGPT operates in 16-bit real mode on an x86 386 processor. The project utilizes the Master Boot Record (MBR)—a 512-byte area of the disk—to house the primary assembly code. This code is responsible for loading an 11.5 KB int8 model from subsequent disk sectors and executing the forward pass to generate a simulated Linux boot log, complete with timestamps and hardware events.
This image represents the primary GitHub repository where the BootGPT project's source code and technical implementation are hosted.
The technical implementation relies on extreme optimization to fit the inference engine into the 510-byte limit of the MBR. The project uses a character-level neural network with one hidden layer and a softmax output. To ensure the model can run on such constrained hardware, the training script quantizes all weights and scales to int8 using power-of-two values. This allows the boot sector to perform the entire forward pass using only `imul` (integer multiplication) and a single `idiv` (integer division) instruction. The runtime handles specific BIOS plumbing, such as using `int 13h` to load weights from the disk and `int 10h` to print the resulting text to the screen.
While the model is not a general-purpose storyteller, it is specifically trained to hallucinate a coherent boot log. The training process involves generating 3 MB of fake `dmesg` (kernel message) data to provide the model with a consistent corpus. Because real boot logs are repetitive, this consistency helps the tiny model produce