(Although this comment discusses an AI-written project, the comment itself is entirely human-written.)

It's probably worth pointing out that, unless you're writing a replacement for GRUB, Multiboot2 is a reasonable alternative to building a UEFI PE32+ image, and it permits you to still call UEFI services: https://www.gnu.org/software/grub/manual/multiboot2/multiboo... https://wiki.osdev.org/Multiboot

I was pleased last week that I could elicit enough of a UEFI environment from Opus 5.5 to be able to build an UEFI application without enormous, un-auditable dependencies like gnu-efi. Claude distilled the relevant parts of the UEFI interface into a 289-line header file. The example application just displays the keyboard and mouse events on a UEFI framebuffer using a vendored pixel font, thus the name "inputmon": http://canonical.org/~kragen/sw/inputmon-3.tar.gz

The Makefile includes a "run" target to boot it under QEMU with the OVMF implementation of UEFI.

(One fairly surprising feature of all this is that, not counting the time Claude was waiting for me, it took the model a total of about 7 minutes to write and test this. These 7 minutes were spread over almost 15 hours because I ran out of tokens and spent time thinking. This might be a good example of a case where an LLM is literally a million times more productive than a human programmer: the task is algorithmically trivial, and it doesn't involve difficult engineering tradeoffs, but involves deep knowledge of a pile of obfuscated recondite interfaces that are publicly documented and widely used by published open-source software. It is something I've wanted to do for 7 million minutes.)

I'm hoping I can get an in-browser demo of inputmon working next week, but that's probably going to involve doing some surgery on Bochs in order to get an amd64-compatible virtual machine that doesn't require multithreading the way QEMU does.

Inputmon does have the enormous, un-auditable dependency of LLVM and Clang, but that's a horror I'm already resigned to. But with clang, lld, and make, you can build the image from source, and with QEMU and OVMF you can run it in a virtual machine.

I think it's a little bit debatable to call UEFI "bare metal" (Boone responds: "this is a fair criticism"), and I'm particularly skeptical of Boone's claim that you can guarantee microsecond response times in the face of SMM, but whatever.

First steps in bare-metal programming using UEFI