Hacker News Reader

Comment by hyperhello | Hacker News Reader

What do LLMs “get right”?

LeetCode questions. They often give the memorized LeetCode answer even if you slightly modify a question so it has a different correct answer.

Lovable, V0, etc. generate landing sites with slick design and easy editability. Startups are going to have such an easy time of product demo and marketing sites.

React is quickly falling to LLMs.

Rust, in my personal experience, not so much.

Whatever code examples got the most votes on StackOverflow.

I seem to use them more and more for repetitive tasks. Where I can write one line and will need to handle a couple more in a similar manner. They work fine.

Also for refactoring they seem to do ok. Things like change this to a function, extract this type etc.

They excel in snippets like “get unique items in this array” or “sort this by property x” kind of stuff where you could easily write or find an answer.

Oh I also like to use them for code review. Not that I’d blindly trust one but you can have another eye to look at your pr (i use claude code for this and love it) and see if you introduced any side effects or missed something.

For anything more complex like having one write some feature from scratch… meh. I haven’t had much luck. Also they seem to fuck up royally in a little complex project if you do not isolate your request like I mentioned above.

I've found them to be pretty darn good at both emacs lisp and matplotlib.

Flutter widget design.

Reply on news.ycombinator.com