Many users affirm the claim that cleaning up LLM code takes as long as writing it, citing hallucinations, unreadable output, and painful debugging.
Most Activity
it's so trash. like it's almost TRYING ITS BEST to produce THE WORST THING POSSIBLE it's impressive
the amount of time i spend cleaning up LLM code is greater or equal to the amount of time it would have taken me to write it myself
@DanielW_Kiwi the new models are reward hacky fuckers
@yacineMTB Weren't you the OG proponent of coding with LLMS? Like you were super early
@Clipart_Bear @DanielW_Kiwi @yacineMTB unironically
@yacineMTB Not even close
@yacineMTB buddy you are the one who convinced me LLM code could do anything and why I gave up my life to become a Sherpa no takes backs
@yacineMTB Stop using the frontier models. Use qwen plus on opencode go, or honestly even qwen 3.6 35b a3b locally is giving me better output than Claude Opus. The frontier models are fucking cooked right now.
@yacineMTB the trick is giving it the existing code as context first and being super specific about what not to touch, still babysitting it but at least it stops rewriting everything lmao
@yacineMTB The funny thing is there are people who are actually paid to review AI output and give it feedback. If they asked me to do RLHF with C++ I would prefer to retire.
@yacineMTB are you gonna write all that cuda yourself?
@yacineMTB tbh it’s because you’re not prompt engineer qualified
@yacineMTB most LLM code online is mediocre, so cleanup often equals a rewrite
@yacineMTB Self report
@yacineMTB 100% working on a side quest to fix this because it's wild
@yacineMTB Same issue here. What I have found useful is to draft out the directory/pseudo-code. Then ask the LLM for very narrow implementations. Anything else and the LLM turns the repo into some of the most unreadable/unintuitive code I have ever dealt with
@yacineMTB Yea =[ I honestly don't know what to do about it. Its just painful
@yacineMTB I thought you posted that software was solved the other day?
@yacineMTB but what do you prefer doing?
@yacineMTB i either don't bother or write myself directly
usually v0 / POC: see if i care, faster is better v1: manual rewrite (i need to maintain that shit so it might as well be mine)
if i can't write it manually it's probably a dumb req anyway v0 → v1 is where i delete half the arch
@todaymare @Clipart_Bear @yacineMTB Yeah, I've noticed this. One advantage of being in NZ is that we are off peak.