Hacker News Reader

Comment by Tade0 | Hacker News Reader

I use LLMs at work. I'm expected to and paid to show up and do that, so I do it.

But I refuse to leave understanding the codebase to machines. I did that the other day as an exception and fortunately my reviewer picked up on things which weren't exactly broken - just redundant and inelegant - e.g. two variables which easily could have been one, passed between several methods.

We have an automatic reviewer as well and there wasn't even a peep from it regarding any of that.

And this is really the crux of the issue - LLMs won't ever create elegant and minimal code, because their style is based on the mean, which in turn is dominated by the work of people who shit out a lot of working, but write-only code.

View on HN

I lost interest in writing elegant code many years before AI. Professional software development doesn't reward it much. Most customers don't care about it. And I don't have much emotional attachment to or ego investment in stuff that I make for an employer, so I don't really care if it's elegant or not as long as it puts money in my bank account.

I used to have more emotional attachment to and ego investment in stuff I made for employers, but professional software development is pretty good at making one stop caring about that stuff. Probably for good reason, since software development is not actually the most important craft at a tech company. Business is more important.

> their style is based on the mean

A base model that has only been pre-trained has output that is based on the mean, but most advances in capability the last three years have come from post-training. The influence of RL dominates every observable characteristic of coding agents today and steers the model towards strategies and practices that are objectively successful for the LLM.

I think the best code for an LLM is different than the best code for a human. Lots of redundancy, lots of unnecessary and overly defensive code is being reinforced by their training because it is successful there.

They still take it to far and do all kinds of stupid shit - if you look at the code you'll see it all the time but it actually doesn't cause many problems. Its much more important to focus steering effort on feature scope and architecture. I still ask it to refactor code from time to time and its pretty good at doing that.

You took the words right out of my mouth. Writing code used to force you to understand it, but now the former doesn’t guarantee the latter. And most attempts to replicate the latter in other ways (more reviews, self-quizzes, etc) seem to be far less effective at producing deep understanding.

Same boat, same reasons.

Skills not exercised degrade as well, if you don't actually write code then your skill at reading and interpretting it degrades, you just become a "looks good" step.

In theory if disciplined enough you could resist that but our brains seem to take the lazy path whenever it is offered, thinking hard and focussing are a skill and one I don't feel is been exercised when using AI.

You're just bad at using LLMs

Reply on news.ycombinator.com