I've seen this explained as LLMs being well - language models. Driver binaries are just a language understood by the machine executing the binary. While its basically unreadable for normal humans, it is just another language for LLMs & they are able to operate on it, changing the result & translating it (e.g. correlating with specs, logs, docs, etc.).

Given the positive results from what I've heard, this might be actually one are where LLMs could prove useful for once.