Ask HN: Should LLM be able to translate C to rust as easy as English to Japanese
Tried that on Google's AI and gpt-oss with few models locally, the answers was not as good and I expect. Any idea on why? Not enough training materials over all?
2 comments
Translating a computer program you can be right or wrong. To write a Rust program that (1) compiles and (2) actually does the job you have to solve constraint problems exactly. If you are translating from English to Japanese you have a wide range of choices, people could argue a lot over whether or not a translation is good, practically one that is "95% correct" by one person's measure might be preferred over a "100% correct" one because people like the style better.
Not necessarily. They're different programming paradigms and the LLM isn't thinking, it's just giving you what is most likely to be the correct translation. But unlike a written language, there's a ton of technical nuance that it has no real awareness of beyond the similarities between tokens.