Are there specific programming fields where large language models (LLMs) really fall short? I'm currently interning as a full stack web developer and have been using LLMs, but I've noticed that relying on them takes away some of the joy of programming. I'm considering shifting my focus to more niche areas, where I might find better job security as companies increasingly adopt LLM workflows. C was my first programming language, and I'm open to exploring lower-level or specialized fields. Any insights on where LLMs might not perform well?
5 Answers
From my experience, LLMs do not handle systems from before the web well at all. If you're tasked with writing code for something like an old Amiga or something in assembly, good luck—it's a struggle for them! They often make up code that doesn't even exist yet.
Honestly, it seems like LLMs struggle across the board! They do well with entry-level code, but once you start tackling production-level tasks with complex codebases, their limitations will become evident. I've seen them mishandle operations like trying to perform math on non-numeric values, which is a pretty basic issue.
So true! It's like they excel with straightforward, new projects but totally flounder when faced with messy legacy systems or intricate code. It's frustrating!
Yeah, my experience aligns with that. I found that while LLMs can assist with simpler tasks, they really struggle when it comes to understanding the logic behind older code or complex systems.
They really seem to struggle with programming contexts that require deep understanding of concurrent or parallel coding. I've heard from some peers in educational sectors that they've had poor results when using LLMs for writing multi-threaded applications.
Exactly, I've faced issues when applying LLMs to parallel tasks. They just don't get the complexity! You really have to know your stuff.
LLMs tend to falter in fields where training data is sparse, such as new embedded systems or niche programming languages. The bigger and more intricate the codebase, the trickier it is for LLMs to keep track of everything, leading to inaccuracies.
In my experience, they perform poorly in computer graphics too, especially with languages like C++. They're not reliable for serious work in that area; it's safer to stick with solid documentation.
Absolutely! I tried to get one to work with some assembly code the other day, and it just couldn't grasp the basics.