Hacker News Reader

Comment by PunchyHamster | Hacker News Reader

> to "for this next step I need a vpn multiplexer written in a language I don't use"

but that acceleration is exactly because you're not good at that language

Can't we reach a compromise where proven track record of good use of LLM by a contributor or a company (eg. Bun) be pre-approved or entertained? Blanket ban on a new technology shouldn't be the default option.

No.

if they had a good track record, the current submission that led to this article damaged it.

i am reminded of this quote: it takes more cleverness to debug code than it takes to write it. if you write code as clever as you can, by definition you are not clever enough to debug it. using LLM makes your code many times more clever than what you could write yourself. which means by the same definition the code is to clever for you to understand or debug it.

I like the new corollary to that rule, which is that if the AI is the best coder in the room and writes code too clever for itself, then no one including the AI can debug the code. Then where does that leave you?

i love it. just a moment of thought makes clear that LLMs are not capable to debug their own code because if they were they would be able to write better code. the LLM code doesn't even need to be clever.

That’s why you don’t use SOTA xhigh models to write your code, so you can use the xhigh model to debug the code.

I kneel to Poe's law.

Certainly not in the case of asking it to do something you'd be slow at because you are unfamiliar. If you are not familiar enough with the system, how are you confident that what the LLM has produced is valid and complete? IMO the people saying LLMs make then 10x faster were either very bad to start with (like me!) or are not properly looking at the results before throwing them over the wall.

And how do you know if that is the case or the person/team using the LLMs is one of the good ones?

So the safest answer is just "no".

This is the crux of the problem. LLMs make me significantly faster at writing code I was mediocre or bad at. But when I use it to write code in domains I have more knowledge in I see design and correctness problems all over the place and actively fix them and it slows down my output.

Speed is seductive.

The bar isn't "this is a known good contributor". Its "this is a known good contributor working in a space they have knowledge in and has a track record of actually checking and thinking about LLM output before submitting it." It's much higher and I don't see how you can approve people on an organization-wide basis.

  > LLMs make me significantly faster at writing code I was mediocre or bad at. But when I use it to write code in domains I have more knowledge in I see design and correctness problems all over the place and actively fix them and it slows down my output.

I think a very similar phenomena is called Gell-Mann amnesia effect: https://en.wikipedia.org/wiki/Michael_Crichton#%22Gell-Mann_...

Why would it be pre-approved ? Code is code, whether it's bad quality LLM code or meatbag code it shouldn't matter.

The entire problem is that before the meatbag code was either not submitted at all (developer knowing they are not competent enough to do the fix) or the volume of it was low.

With LLM people not competent to even review, let alone write are emboldened to just throw shit on the wall at rapid pace. So the wall is entirely covered by the shit

Reply on news.ycombinator.com