Comment by mattmanser | Hacker News Reader

Comment by mattmanser | Hacker News Reader

Parent Full thread

Copilot? I personally found Copilot pretty useless. And very distracting trying to constantly incorrectly guess what I was doing.

I do use chatgpt/claude for quickly creating first draftw of frontend pages as I find it quicker, but I have to do extensive edits to get it right. But anything non-boiler plate I am faster writing it myself.

I have also quickly found it is a stochastic parrot but that's ok for boilerplate stuff. It is useless at anything complicated and quickly starts hallucinating methods, parameters and functions when you point out problems with the code.

I am beginning to believe that anything niche, where it can't steal from its training data, and it will always be utterly useless.

I am curious how you feel it's so important to 100+ engineers, how do you use it that makes it effective for you?

Thank you! Maybe there is just a divide between engineers working on deep complex topics and those trying to setup their 17th Django app.

LLMs are a totally useful tool! They definitely help me write all the bash scripts I consistently forget.

But the deep stuff where the value lies? It starts to fall apart quickly.

You put this better than I could have. I am really impressed by LLMs' ability to respond to a prompt with code which compiles and runs. That's a hugely impressive result. But it doesn't mean the technorapture is just around the corner. There's a vast gulf between what LLMs can do currently and what skilled people do, and it's not "obvious" that there's some path from here to a world in which LLMs become some kind of superintelligent engineer-gods. It could be physically impossible, for all we know. That's why the pseudoreligious twaddle gets under my skin so much--it's the naive assumption that progress is linear, instead of a jumpy "fits and starts" process of accidental discoveries building on one another (or encountering dead ends). There's no way to know whether these things can be improved enough to be truly useful for real work and pretending it's "right around the corner" won't necessarily make it happen.

Reply on news.ycombinator.com