Limity programování s asistencí AI
Nedávno jsem psal o tom, jak se díky nástupu jazykových modelů postupně mění role vývojáře. Jon Gjengset, autor knihy Rust for Rustaceans, nabízí pohled na jeden z limitů, na které umělá intelince, resp. algoritmy strojového učení, naráží (viz jeho lednové Q&A (zvýraznění níže je moje)):
I think in particular where [machine learning] works really well is when you have either pattern matching or pattern recognition, where you have a lot of data to source those patterns from. So I don’t generally think of machine learning as being smart. I think of it as being really good at spotting patterns that people have seen before, or replicating patterns that have been seen before. And that could include combining patterns in—let’s call them—novel ways, but at least combining them in some way that might appear novel.
I think there are a number of maybe surprising areas where this works really well. One of them is code generation. I don’t use it a lot for my coding, and there are a couple of reasons for that, but one that I’ve touched on in the past is that the code I tend to write now, and the code I was writing at AWS, and the code I was writing during my PhD, was not very standard, for lack of a better word. Like, most of the code—it is unclear if anyone had written quite that code before. That’s not the case for a lot of software development.
