Hacker News Reader

HNHacker News

Top New Best Ask Show Jobs

Comment by jayd16 | Hacker News Reader

Parent Full thread

jayd16·

Unreliable tools are utterly exhausting.

> not much worse than a junior dev and 100x faster.

Is there a greater hell than this!?

View on HN

If the old metric is right, that it is ten times harder to debug code than to write it, having something that writes buggy code 100x faster than you can understand it is a problem.

Especially given that you can ask an LLM to optimise code and on multiple runs it can not tell if it's is improving or degenerating.

At least with a junior dev, I can teach them how to do it better next time. Not so much with generative "AI".

Not totally. But you might be surprised at the things you can do. Cursor has some template-like files where you can basically teach the AI “when we do X, do it this way.” Or you can change the global prompt to add the things it should keep in mind when working with you.

If you actually take the time to tell it “hey, don’t do it this way,” it can definitely do it differently the next time.

On top of that, is anyone training models on their own codebase, and noting to AI which patterns are best practice and which aren’t?

There are a ton of ways to make it better than the baseline copilot experience

> a junior dev and 100x faster. > Is there a greater hell than this!?

Yes — junior management using LLMs and 100x more cocksure.

That's 100x more bugs to fix. Moreover, increasingly complex models produce bugs that are increasingly hard to spot and fix.

I am of the firm belief that unreliable help is worse than no help at all. LLMs are unreliable help, therefore they are useless to me.

LLMs are the "move fast and break things" of AI.

Worse than an intern and 1000x faster?

"It makes lots of mistakes, but at least it makes them quickly!"

Reply on news.ycombinator.com