Post #4313389

I know it's nothing new, but I'm still gonna say it: why do people use LLMs for coding? I have access to LLMs at work, so from time to time I test them (different models, different problems), but almost every time they return garbage code. - Recently I asked it to write code and avoid a specific pattern that was introducing a security vulnerability. The LLM insisting on the pattern and re-introducing it multiple times, even if reminded that it should not. - In other case, it suggested "performance improvements", on "you're wrong" it run the benchmarks and indeed, it was wrong. - In other case, it suggested a "performance improvement" of caching a value instead of repeated function calls. When benchmarked it changed nothing, the compiler already seemed to figured it out itself. - In many cases it introduced some useless local optimizations that only made code uglier and more complicated without any benefit. and neither of the cases was some subtle edge cases, it was really simple stuff. It's basically a tool that copies and pastes the answers from StackOverflow until the code compiles and tests pass, with some ability to adapt them to local context, but that's basically it. Yes, you can solve many problems like this. Yes, this is what many people often did. But it has nothing to do with LLM having any clue what it is doing with your code, let's be clear on this.