Language is language. To an LLM, English is as good as Java is as good as machine code to train on. I like to imagine if we suddenly uncovered a library of books left over from ancient aliens, we could train an LLM on it (as long as the symbols themselves are legible), and it would generate stories in the alien language that would sound correct to the aliens, even though the alien world and alien life are completely unknown and incomprehensible to us.
Comment on Why don't these code-writing AIs just output straight up machine code?
some_guy@lemmy.sdf.org 3 days ago
No one is training LLMs on machine code. This is sorta silly, really.
TauZero@mander.xyz 2 days ago
aubeynarf@lemmynsfw.com 2 days ago
not necessarily, just as interpreting assembly to understand intent is harder than interpreting “resultRows.map(r -> r.firstName)”, additional structure/grammar/semantics are footholds that allow the model to form patterns at a higher level of abstraction
TauZero@mander.xyz 2 days ago
Only because it’s English and the model is already trained on a large corpus of English text, so it has some idea of what a “table row” is for example. It could learn the concept from reading assembly code from scratch, it would just take longer. Hell, even Lego bricks can be trained on! avalovelace1.github.io/LegoGPT/
Our system tokenizes a LEGO design into a sequence of text tokens, ordered in a raster-scan manner from bottom to top. … At inference time, LegoGPT generates LEGO designs incrementally by predicting one brick at a time given a text prompt.
Knock_Knock_Lemmy_In@lemmy.world 2 days ago
Decompiling an executable into human readable code could be useful. But you would probably train on the opcodes, not the machine code.
naught101@lemmy.world 3 days ago
I think on top of this, the question has an incorrect implicit assumption - that LLMs understand what they produce (this would be necessary for them to produce code in languages other than what they’re trained on).
LLMs don’t product intelligent output. They produce plausible strings of symbols, based on what is common in a given context. That can look intelligent only in so far as the training dataset contains intelligently produced material.