pull down to refresh
10 sats \ 0 replies \ @mrsu 17 Jun \ on: My bad experiences using AI as a physicist science
As a programmer and researcher, I find LLMs are useful for:
- generating simple boilerplate code (that I already know how to create myself)
- simple error debugging
- discovery of new libraries for solving a particular problem
- "smart" google search (to give your very specific question an answer faster than trawling through forums)
- as a light personal teaching assistant, e.g. for learning the fundamentals of a new programming language
- as a proof reader for my notes to make sure my research ideas are on track
What I have found LLMs are not useful for:
- writing whole programs (vibe coding be damned)
- writing high quality code that works
- debugging more than a simple error
- giving you a better answer than the man pages (RTFM is still relevant - even if it takes more time)
- giving accurate information you can rely on.
- writing in general (they're too verbose)
I used to rely on LLMs quite heavily, but I found they usually just ended up wasting more time than they saved. Quite often they would lead me down a deep rabbit hole that went nowhere, and to get back on track, I'd end up needing to research the topic manually myself anyway.
Today, I use them sparingly. I'll only use them when I know an LLM will give me a good answer, otherwise I don't bother. I've mostly reverted back to Google search and reading man pages for my work.
They have their uses, its just about understanding how and when to use them. They aren't a replacement for traditional cognitive tasks such as researching or programming. Rather, they're more like a helpful junior assistant that require constant supervision and prompting to get any use out of them.