An AI small enough to fit in a PC's boot sector

What did the maker build?

When a PC starts, it first runs a tiny program of 512 bytes, the boot sector (the first sector of the start-up disk). Normally that program loads Windows or Linux.

This project, called bootgpt, uses those 512 bytes to run a neural network (a program that learns from examples) instead. The README on GitHub says the machine wakes up and “dreams” a Linux start-up log: times, hardware names and messages for parts that never existed.

Why is it so small?

The README gives the sizes.

The training script shows the rest. The network has an embedding (a small list of numbers for each character), one hidden layer of 96 units, and 64 possible characters. Add it up and it is about 11,000 numbers.

It was trained on 3 MB of fake start-up logs that the author generated. The README says real logs repeat a lot, and that repetition is what makes a tiny model look sensible.

What does it write?

This is a sample from the README, so you can judge the quality yourself.

[  1.883084] NET: Registered protocol family 16
[  2.283888] pci 0000:01:01.4: [1b36:0a35] ryndrde c4128 6422 Vveine

The first line looks real. The second starts well and then turns into nonsense words. That is what you expect from a model that only sees 8 characters at a time.

Why should you care?

Today’s chatbots have billions of numbers. This one has about eleven thousand, and it still does the same job: it guesses the next character, over and over.

So the idea of a language model is small. What changes with size is how much it can remember and how sensible it sounds. A model this small can only learn one narrow thing, and here that thing is a boot log.

What’s the catch?

• It does one trick. It cannot answer questions or hold a conversation.
• The numbers are the author’s own, from the README. We did not build or run it.
• You need two tools, both listed in the README. NASM builds it. QEMU (a program that pretends to be a PC) runs it.

Try it

On paper, write ten lines of a pretend log, such as [ 0.100] disk ready. Cover all but the last 8 characters of a line. Guess the next character. Then guess again. You are doing what bootgpt does, with a much better memory.

If you are comfortable with a terminal, the README lists the steps to build the image and run it in QEMU. Start by reading train.py.

Sources

• bootgpt, README and train.py, kerlenton, GitHub (no publication date given; read on 1 November 2026)