Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

LLMs can get "brain rot"

llm-brain-rot.github.io

Post reply on HN
Print view View original

221–230 of 310 posts

Re: LLMs can get "brain rot"

#221

Post by mikeiz404 »

standardly wrote: post #93
avazhi wrote: post #47

“Studying “Brain Rot” for LLMs isn’t just a catchy metaphor—it reframes data curation as cognitive hygiene for AI, guiding how we source, filter, and maintain training corpora so deployed systems stay sharp, reliable, and aligned over time.” An LLM-written line if I’ve ever seen one. Looks like the authors have their own brainrot to contend with.

That is indeed an LLM-written sentence — not only does it employ an em dash, but also lists objects in a series — twice within the same sentence — typical LLM behavior that renders its output conspicuous, obvious, and readily apparent to HN readers.

Ah now that's the kind of authentically human response I was hoping for!

(It's a joke: The parent uses the same writing style they described as being indicative of LLMs)

Re: LLMs can get "brain rot"

#223

Post by Rileyen »

After reading this, I just felt like everyone already knows the data is a mess, but no one really cares. We feed the models a bunch of junk, then act surprised when they start getting dumber. Honestly, did we even need a study to figure that out?

Re: LLMs can get "brain rot"

#224

Post by ramon156 »

yieldcrv wrote: post #220
nazgul17 wrote: post #219

Earlier quoted context omitted.

They also tried to heal the damage, to partial avail. Besides, it's science: you need to test your hypotheses empirically. Also, to draw attention to the issue among researchers, performing a study and sharing your results is possibly the best way.

I don’t understand, so this is just about training an LLM with bad data and just having a bad LLM? just use a different model? dont train it with bad data and just start a new session if your RAG muffins went off the rails? what am I missing here

Do you know the conceot of brain rot? The gist here is that if you train on bad data (if you fuel your brain with bad information) it becomes bad

Re: LLMs can get "brain rot"

#225

Post by ErroneousBosh »

turtletontine wrote: post #101
standardly wrote: post #93

Earlier quoted context omitted.

That is indeed an LLM-written sentence — not only does it employ an em dash, but also lists objects in a series — twice within the same sentence — typical LLM behavior that renders its output conspicuous, obvious, and readily apparent to HN readers.

I think this article has already made the rounds here, but I still think about it. I love using em dashes! It really makes me sad that I need to avoid them now to sound human https://bassi.li/articles/i-miss-using-em-dashes

I use them too, and there's not a trace of artificial intelligence in my posts - it's good old-fashioned analogue stupidity all through.

Re: LLMs can get "brain rot"

#226

Post by mikeiz404 »

> "brain rot", "Thought-skipping", "primary lesion", "Cognitive Declines", ...

In general using these medical/biological metaphors doesn't seem like a good idea in things like computer science research papers and similar.

Their use forces many inaccurate comparisons (when compared in detail) and they engender human qualities to what are already forgotten to be just computer models. I get this may be done with a slight tongue-in-cheek but with research papers there is also the risk that these terms start to be adopted. And undoing that would be a much taller order in either the research community or general media.

Maybe I am just yelling at clouds.

Re: LLMs can get "brain rot"

#228

Post by close04 »

TeMPOraL wrote: post #217
lxgr wrote: post #210

Earlier quoted context omitted.

The solution is clear: Unicode needs cryptographically signed dashes and whitespace characters.

Tied to what? Show us a way to create a provably, cryptographically integrity-preserving chain from a person's thoughts to those thoughts expressed in a digital medium, and you may just get both the Nobel prize and a trial for crimes against humanity, for the same thing.

Why don't you come say that to my face?

Re: LLMs can get "brain rot"

#229

Post by f_devd »

gaogao wrote: post #67
f_devd wrote: post #56

Earlier quoted context omitted.

I do not think this is the case, there has been some research into brainrot videos for children[0], and it doesn't seem to trend positively. I would argue anything 'constructed' enough will not classify as far on the brainrot spectrum. [0]: https://www.forbes.com/sites/traversmark/2024/05/17/why-kids...

Yeah, I don't think surrealism or constructed is good in the early data mix, but as part of mid or post-training seems generally reasonable. But also, this is one of those cases where anthropomorphizing the model probably doesn't work, since a major negative effect of Cocomelon is kids only wanting to watch Cocomelon, while for large model training, it doesn't have much choice in the training data distribution.

I would a agree a careful and very small amount of above brainrot in post-training could improve certain metrics, if the main dataset didn't contain any. But given how much data current LLMs consume and how much is being produced and put back into the cycle I doubt it will miss be missed
Post reply on HN