Hacker News Reader

Comment by al_borland | Hacker News Reader

From what I’ve seen, LLMs are good at making stuff that has already been made and posted to GitHub a thousand times before. At my job we’re constantly asked to do things that really haven’t been done before, at least not by people sharing source code, so the LLMs suck at most of it.

LLMs make for great tech demos, but when it comes to writing code for production that actually does something new and useful, it hasn’t impressed me at all.

View on HN

You have to pair program with them. They don't work well at all when fully autonomous.

But they do work for pair programming. Which explains a lot of the tech layoffs we've been seeing.

Ain't nothing new under the sun.

That’s what I added the caveat about it being open sourced. I’m sure what I’m doing has been solved many times by many companies, but it’s not information they share publicly, just like my code isn’t shared publicly. There is also the issue of integrating with existing legacy systems, which may include bespoke internal tools.

Maybe the thing has been done in general, but not in a way that’s useful for me. That’s why it looks good in tech demos. If I ask AI to write what I need, it will give me an answer, but it won’t actually work and integrate in the ways I need for production. The last time I tried it gave me 70 lines of code, the real end result was thousands. The AI version would look cool in a demo though.

Reply on news.ycombinator.com