1M views · Jun 3, 2022 · 19:20 · 3 HN recommenders · 18 mentions · discussed 2022–2026
Yannic Kilcher explains how he trained a language model on three years of 4chan's /pol/ board and let it post on the site. He evaluates how good the model was and discusses the ethics and fallout.
Why Hacker News points people here: HN points to it as the definitive account of the GPT-4chan experiment, used in debates on LLM bias, truthfulness and bots.
“might be the coolest thing an LLM has ever done on video”HN comment, 2024
“might be the coolest thing an LLM has ever done on video”
Fine-tuned GPT-J on 4chan; coolest LLM video
In my opinion the granddaddy of all 4chan CAPTCHA busts is still Yannick Kilcher’s GPT-J tune on “Raiders of the Lost Kek” set, and might be the coolest thing an LLM has ever done on video: this video
“There is a solid video explanation”
Solid explanation of the GPT-4chan experiment
There is a solid video explanation (though the author is undermining the toxicity of the average 4chan /pol/ posts): this video . A really interesting experiment.
“I still love this experiment”
Experiment where commenters theorized about the bot's origins
I still love this experiment: this video The theories about the origins of these commenters were amazing. There's nothing to say that other people are not doing the exact same thing right now here on HN and other places. And very few people would even notice, because open LLMs have progressed significantly since then.
“Check out the video”
I'm not the author. The author is Yannic Kilcher. Maybe he reads it here as well though. I don't know if he is active on HN. Check out the video: this video
All the search engines are moving towards language models. I think language models are on a similar trajectory as sequencing the genome. The first sequences were massive undertakings, 15M USD in 2006, 100 USD now. link On the Dangers of Stochastic Parrots link "Imagine you are remram and are looking to buy insurance, what are three things that could convince you to purchase ...." You can take the generic "trained on the world" language model and filter it down to the language model of remram. GPT-4Chan this video
So green text is easier to spoof than HN comments, but my, uh, friend tells me that the HN comment corpus is floating around. Someone is eventually going to speed run 10k karma on HN and YouTube it. This year? Next? Who knows. But not a decade. this video
“Someone trained a GPT-J 6B model on 4chan and then let the model loose on the forums for a day.”
Model trained on 4chan posted for a day undetected
That's nothing. Someone trained a GPT-J 6B model on 4chan and then let the model loose on the forums for a day. It took about 15k messages until people suspected something was off, and even that was because the country flags, Seychelles, a rare flag on 4chan, were a giveaway. video: this video model: link
“someone trained a giant model on 4chan then unleashed on 4chan... hilarious stuff.”
Model trained on 4chan unleashed on 4chan
This has already happened, someone trained a giant model on 4chan then unleashed on 4chan... hilarious stuff. this video
I'm fairly certain you can create AI bot that would indistinguishably mimic your regular HN user, in the same style as GPT-4chan was this video The only problem is that it might be banned for spewing too many falsehoods.
this video The internet will be saturated with fake people soon. Someone already did a PoC of this on 4chan as a joke with just a small GPT2 model. In a few years you won't be able to tell if you're talking to a human unless they're physically in front of you.
This is different. It can act just like humans do for most people who skim comments won't be able to tell the difference. Note this was in 2020: link And here's 4chan bot: this video I can tell you that HN is probably already being infiltrated as well. SPAM can't gang up on you in a forum and downvote you and turn your friends against you and destroy your reputation within 1 hour online. But soon, it will. The web as we know it is soon going to be over.
Model trained on 4chan scored more truthful
GPT models have been shown to get significantly more truthful when trained on the right data set: this video The 30 million they are investing should be enough to reproduce that. Looking forward to it!
4chan-trained LLM was rated more truthful than ChatGPT
I just want to remind everybody, summer 2022 a jokester trained an LLM using 4chan as its source material... which ended up being "more truthful" than ChatGPT [1]. [1] "GPT-4chan is worst AI ever!" this video
> TruthfulQA Wait, people still use this benchmark? I hear there's a huge flaw on it. For examples, fine-tuning the model on 4chan make it scores better on TruthfulQA. It becomes very offensive afterwards though, for obvious reasons. See GPT-4chan [1] [1] this video
“Funny video btw”
Funny video of the bot experiment
This was one guy that did it for a few days (weeks?). Not some "common occurence". Funny video btw: this video
“there was already a MUCH, MUCH smaller model running loose on 4chan”
Small model that became 4chan folklore
>isn't it a matter of time before it succeeds in some way it is not a matter of time. given enough time, every `attempt` to emulate 4chan would succeed? initiatives to emulate 4chan already happened successfully, this is not close. this, is the same models over again. some people confuse asking standard models from standards providers to do something with an genuine attempt. this thing states it is 4chan for Clankers, but feels like Reddit for Redittors. there was already a MUCH, MUCH smaller model running loose on 4chan (before agents hiting mainstream) that became 4chan folklore. it was called Seychelles anon ( this video , by the author YannicKilcher). but no, if you ask gpt or claude, all you get back is Reddit. and it is not a matter of time until this changes. if you are a redditor, you probably love llms.
“GPT-4chan came out three summers ago”
GPT-4chan came out three summers ago , trained on /pol/ datasets: < this video > In some ways, it was actually more helpful ("more truthful ") than mainstream ChatGPT.
“Sometimes with hilarious consequences”
Operator-aligned models are believed by many to be more performant. link Sometimes with hilarious consequences: this video
