Hacker News Reader

Comment by xanderlewis | Hacker News Reader

The main reason (what we’re now calling) AI (specifically, LLMs) is so popular here is that they happen to be a somewhat useful tool for programmers. Other than that, they’re not really that great. Unless you want slop: be it writing, art, music, or whatever. Use it for inspiration — nothing else. And certainly don’t trust it.

View on HN

> they happen to be a somewhat useful tool for programmers

Outside the junior/mid level though you've probably mastered a combination of skill and research mastery such that an LLM would slow you down.

It's pretty hard to beat an LLM for filling in common boilerplate or autocompleting relatively obvious lines of code that can be checked in a second or two, no matter how much experience you have.

Adding to that, there's definitely techniques to make it autocomplete less trivial lines of code as well. Using comments and clear function names and sane abstractions to nudge it in the right direction helps a lot.

Hard to describe, but Andreas Kling's JIT compiler youtube series is a good demonstration of this technique.

I'm dumbfounded that anyone technical can feel this way. The more I use LLM tools, the more it feels like a super power.

Since you talking technical, is there some data to back that up or is it just vibes?

Number of Stackoverflow visitors going down by a lot, come to mind

Seems it was 14% when gpt4 came out https://stackoverflow.blog/2023/08/08/insights-into-stack-ov...

so noticeable, but not a lot, (though don't know the current data)

I only have personal data. Things that would have taken me a long time before take me much less time now.

As an LLM skeptic, I will say it’s been nice for learning new languages and frameworks for a side project (e.g. learning js/react as a backend/data guy). Things I’d probably not have bothered to learn because it’s just too much work and I’ve got enough value in my current skills. I don’t find it particularly helpful in my primary languages/frameworks though.

I also use it to generate some discussion on ideas. For example, the other day I asked if to consider pro/cons of technical approach A vs B. I don’t have to trust the output, but it’s usually good enough at helping me sketch out the problem space at a high level.

AI autocomplete can often do things faster than I could physically type the code, and it can do trivial but very uncommon things that I haven't encountered enough to actually learn, like sed/awk/etc syntax.

Reply on news.ycombinator.com