If it is too big just zip it and feed chunks of the resulting zipfile to the AI. AI can do anything, right?

It will work if you convert zip chunks to base64 first, and use a large enough training set.

Should work fine if the training set is a memory-safe, full-featured, monolithoc, Unix-like operating system kernel originally written by a certain Linus Torvalds.

Can't wait to read the first incendiary linus style rant being generated by the AI as a result.

"As a large languge model, I can't answer your patch merge request you absolute fucking moron"

Easy, triple the hardware requirements and don't talk to any hardware because if you do you'll have to mess with buffers in a non approved way.