... until you realize that the LLM-generated code doesn't even compile, or you need a PhD to write all the prompts needed to have a prototype instead of the real thing.
why doens't someone train an LLM on predicting source code given a sequence of input machine code tokens? that is so obvious. why does nobody do it?
or maybe its not even neccesary, and doing something akin to fuzzing syscalls 'but smartly' probably yields more results.