66 comments

[ 4.4 ms ] story [ 86.7 ms ] thread
Isn't this just garbage in garbage out with an attention grabbing title?
> continual exposure to junk web text induces lasting cognitive decline in large language models (LLMs).

TLDR: If your data set is junk, your trained model/weights will probably be junk too.

AIs need supervision, just like regular people... /s
Another analogy to help us understand that LLMs are a useful part of what people do but are wildly misconstrued as the whole story
making a model worse is very easy.
Ah yes, something the local LLM fine tuning community figured out how to do in creative ways as soon as llama 1 released. I'm glad it has a name.
If only I got money every time my LLM kept looping answers and telling stuff I didn't even need. Just recently, I was stuck with LLM answers, all while it wouldn't even detect simple syntax errors...
This paper makes me wonder the long lasting effects of the current media consumption patterns by the alpha-gen kids.
“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.

(comment deleted)
I recently saw someone on HN comment about LLMs using “training” in quotes but no quotes for thinking or reasoning.

Making my (totally rad fwiw) Fiero look like a Ferrari does not make it a Ferrari.

What is actually up with the "it's not just X, it's Y" cliche from LLMs? Supposedly these things are trained on all of the text on the internet yet this is not a phrasing I read pretty much anywhere, ever, outside of LLM content. Where are they getting this from?
It's probably getting amplified by the RLHF stage because the earlier models didn't do that.

But that just shifts the question to "what kind of reviewer actually likes 'it's not just X' cliche?" I have no idea.

I think using large language models really accelerates mental atrophy. It's like when you use an input method for a long time, it automatically completes words for you, and then one day when you pick up a pen to write, you find you can't remember how to spell the words. However, the main point in the article is that we need to feed high-quality data to large language models. This view is actually a consensus, isn't it? Many agent startups are striving to feed high-quality domain-specific knowledge and workflows to large models.
HR people have been speaking that way long before LLMs.

Did you already update and align your OKR’s? Is your career accelerating from 360 degree peer review, continuous improvement, competency management, and excellence in execution? Do you review your goals daily, with regular 1-on-1 discussions with your Manager?

It is sad people study "brain rot" for LLMs but not for humans. If people were more engaged in cognitive hygiene for humans, many of the social media platforms would be very sane.
I wish I had your confidence in "detecting" LLM sentences. All I can do for now is get a very vague "intuition" as to whether a sentence is LLM-generated. We know how intuitions are not always reliable.
This is pretty clearly an LLM-written sentence, but the list structure and even the em dashes are red herrings.

What qualifies this as an LLM sentence is that it makes a mildly insightful observation, indeed an inference, a sort of first-year-student level of analysis that puts a nice bow on the train of thought yet doesn't really offer anything novel. It doesn't add anything; it's just semantic boilerplate that also happens to follow a predictable style.

I think it's funny/logical how research suggests LLM use makes the user—who is writing more content for the LLM to consume, of course—less intelligent, which makes the system get less intelligent over time.

Sugar, alcohol, cigarettes, and LLMs.

Im curious where all you top commenters were 5 years ago when grammarly was a product used by most professional writers.

If you weren't as incensed then, it's almost like your outrage and compulsion to post this on every hn thread is completely baseless.

Perhaps because it didn't stick out like a sore thumb? Or because it became so prevalent they observe the exact same tics in every other article they read nowadays?
Brain rot texts seems reasonably harmful, but brain rot videos are often surreal and semantically dense in a way that probably improves performance (such as discussed on this German brain rot analysis https://www.youtube.com/watch?v=-mJENuEN_rs&t=37s). For example, Švankmajer is basically proto-brainrot, but is also the sort of thing you'd watch in a museum and think about.

Basically, I think the brain rot aspect might be a bit of terminology distraction here, when it seems what they're measuring is whether it's a puff piece or dense.

This is a potential moat for the big early players in a pre-atomic steal sort of way as any future players won’t have a non-AI-slop/dead internet to train new models on.
I encourage everyone with even a slight interest in the subject to download a random sample of Common Crawl (the chunks are ~100MB) and see for yourself what is being used for training data.

https://data.commoncrawl.org/crawl-data/CC-MAIN-2025-38/segm...

I spotted here a large number of things that it would be unwise to repeat here. But I assume the data cleaning process removes such content before pretraining? ;)

Although I have to wonder. I played with some of the base/text Llama models, and got very disturbing output from them. So there's not that much cleaning going on.

Karpathy made a point recently that the random Common Crawl sample is complete junk, and that something like an WSJ article is extremely rare in it, and it's a miracle the models can learn anything at all.
My son just sent me an instagram reel that explained how cats work internally, but it was a joke, showing the "purr center" and "knocking things off tables" organ. It was presented completely seriously in a way that any human would realize was just supposed to be funny. My first thought was that some LLM is training on this video right now.
is that why chatGPT always tells me "6 7 lol"? ;)
" 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."

Is this slop?

duh! isn't that obvious. is this some students wanted a project with pretty graphs on writing experience?! I am not trying to be cynical or anything. just questioning the obvious thing here.
My Goodness, looks like Computer 'Science' is a complete euphemism now.
It's turning into a social science.
Our metaphorical / analogical muscle is too well developed. Maybe there is a drug we can take to reduce how much we lean into it.

If you look at two random patterns of characters and both contain 6s you could say they are similar (because you’re ignoring that the similarity is less than 0.01%). That’s how comparing LLMs to brains feels like. Like roller skates to a cruise ship. They both let you get around.

I don't understand why people have a hard time understanding 'garbage in, garbage out'. If you train your model on junk, then you will have a junk model.
I suspect this is “the computer is always right”.

A lot of people think computers have better answers than people.

AI is just another type of computer. It knows a lot of things and sounds confident. Why wouldn’t it be right?

"Brain rot" is just the new term for "slang that old people don't understand".

"Cool" and "for real" are no different than "rizz" and "no cap". You spoke "brain rot" once, and "cringed" when your parents didn't understand. The cycle repeats.

Based on my hobby of collecting old gaming magazines, I think 1980s UK had the best slang.