IMO it doesn't make sense for LLMs to write machine code, because abstraction allows them to have more information density. As context size is a big limitation right now, cluttering it with less dense information seems like a bad idea.
What actually might could make sense in the future would be super-dense but abstraction-able programming language made for models. Then you'd train the model to work with it and you could have a way to render a human-readable version of the program for introspection.
It preferably come with superior guard rails, so strict static typing, borrow checker if not gc, perhaps ability to state proofs, etc.
IMO it doesn't make sense for LLMs to write machine code, because abstraction allows them to have more information density. As context size is a big limitation right now, cluttering it with less dense information seems like a bad idea.
What actually might could make sense in the future would be super-dense but abstraction-able programming language made for models. Then you'd train the model to work with it and you could have a way to render a human-readable version of the program for introspection.
It preferably come with superior guard rails, so strict static typing, borrow checker if not gc, perhaps ability to state proofs, etc.
Probably a long time, as machine code is verbose and thus destroys the context both when reading and writing, as well as making it more expensive.
Direct 1s and 0s