Hacker News Reader

Comment by sph | Hacker News Reader

I want to believe this is witty sarcasm, given the topic, but you never know these days.

No, it's been my experience asking LLMs to explain complicated code to me. They can do a pretty good job of it. You can also ask them to write a bunch of tests and then rip out the old implementation and fix it with a new one. LLMs aren't perfect but they are really good at understanding and writing code.

It depends on what one's idea of an implementation is and what one's expectations are.

If an implementation is "a black box that works," absolutely, LLMs are fine. They can create complex programs in full.

But if one expects a tight implementation... it's a horror. LLMs vastly overengineer code (I believe they're intentionally designed to do so). I really struggle to make LLMs generate lean code.

In the best case, they generate bloated designs (in terms of complexity, not necessarily performance), but in the worst case, they generate gaps in the specification. And even in the first case - cognitive load is a problem also for LLMs, not just for humans (although obviously at a larger scale).

I think LLMs are essentially modern compilers, with similar problems, but designed more like "deoptimizing compilers" than optimizing ones. :)

In this subthread we’re talking about being paid to untangle LLM slop, and you come in to say “LLMs are good for that.” Come on man, be serious.

It's right though.

There exists no evidence that manually trying to fix or rewrite the system is going to be more efficient today.

And if AI improves even half as fast as it did over the last year, all bets are off for what 2028 will look like.

Reply on news.ycombinator.com