LockerGnomeTech news and nostalgia for today's technology professionals
Subscribe

Search

Softwarebootgptboot sectorassemblyneural networksretro computingopen source

A free 512-byte boot-sector program, bootgpt, boots a PC straight into a neural network that dreams fake Linux logs

Semyon Usachev's bootgpt packs a whole neural network's runtime into the 510 bytes of a PC's master boot record. It loads an 11.5 KB model from the next 23 sectors and prints a Linux boot log for hardware that was never there.

Chris, built from glowing text-mode block shading, leans close to an old beige computer monitor full of scrolling lines
Illustration: LockerGnome

The output looks right from across the room. Timestamps in brackets, PCI addresses, memory ranges, ACPI interrupt links, the line about registering protocol family 16. Lean in and it falls apart into words like “ryndrde” and “usbbore.” The README calls it a machine that “wakes up and dreams a Linux boot log: timestamps, PCI bars, ACPI events for hardware that never existed.”

The model is small on purpose. It’s a character-level net with eight characters of context and one hidden layer, trained on 3 MB of fake boot logs that a script generates. The generator keeps one kernel version and one disk per session, “because real boot logs repeat themselves and that repetition is what makes a tiny model look sane.” It does one job, and that job is pretending to be the machine it’s running on.

The developer, Semyon Usachev, posts on X as krlntn, and the repo links his build thread.

Running it takes four commands on a Mac: install nasm and QEMU with Homebrew, set up a Python environment with numpy, then ./build.sh and ./run.sh. The disk image it makes “is a normal boot image,” and the README says you can write it to a USB stick and boot any BIOS machine with it. “It only reads its own sectors.” Any keypress reboots.

Boot sectors have a long line of these stunts, and the README credits them. Oscar Toledo fit chess, a whole operating system and a BASIC interpreter into 512 bytes. The closest relative is sectorllm, which boots a real five-layer llama2 transformer from about 1.2 KB of real-mode assembly. bootgpt went the other way, with a smaller engine and a model that knows exactly one thing.

The README sets the bar plainly: with eight characters of context and one hidden layer, the sample log “is roughly the quality you get.” What it doesn’t say is how much better the dream gets with a sixteenth character, or whether 510 bytes has any room left to find out.

More Software

Every Software story →