GPT-4chan | Hacker News Reader

GPT-4chan

gpt-4chan.com

236 pointsalbertzeyer148 comments

Seeing this I initially let out a laugh until I scrolled down and saw that they trained this using /pol/. Even as cynical as I am I don't think the world deserves a infinite racism machine.

Why not train it using a more whimsical (although still offensive) board? You could probably automate all videogame discussion forevermore using /v/.

Probably because /pol/ has the most activity and therefore training data.

https://4stats.io/

Maybe because it was the dataset the author found

I'd rather /g/ for the infinite gentoo memes, but seriously there are multiple boards more interesting than the stormfront containment board

There's no such thing as a "containment" board. They don't keep anything or anyone contained.

Containment boards don't work, but my recollection is that /pol/ was nevertheless created with the intention of being a containment board. /n/ and /new/ had already been banned for racism, but it metastasized across the rest of the site. /pol/ was created after banning boards for racism failed.

Someone should make a GPT from the SCP wiki.

/mlp/ is the only actual containment board.

More like "magnet"

A guy from /g/'s daily programming thread started collecting posts for that exact purpose. Nothing came of it but it's still a hilarious snapshot of that general. Sadly it's dead now.

https://twitter.com/dpttxt

> trained this using /pol/

I wasn't expecting the g in GPT to stand for gestapo

Underrated post here

The pages are littered with warnings that it's vile, so not sure why the top comment is basically repeating that fact without adding any information.

It may be just a question of days until somebody hooks it up back to /r/pol. I bet quality of discussion won't decrease though.

> why not...

Have you ever heard about lulz? The internet hate machine? A true 4chan native will train a 4chan bot with /pol/ and probably with /pol/ only.

For years now, there have been seemingly crazy people on 4chan who insist that systems like this are already employed on 4chan to guide or simply disrupt conversations. Maybe they were crazy to think that a few years ago (or maybe not) but I expect they definitely feel vindicated now.

Related: "Is The Government Spying On Schizophrenics Enough?" https://www.youtube.com/watch?v=FzoXQKumgCw

>For years now, there have been seemingly crazy people on 4chan who insist that systems like this are already employed on 4chan to guide or simply disrupt conversations.

I'd say that probably from about 2009 and certainly from the onset of the Trump candidacy, it would be absurd to suggest there wasn't at least some degree of astroturfing occurring on /b/ and /pol/ (aka /new/ depending on what time frame we're talking about) at a minimum.

It's important to remember that 4chan was a lot more influential in the past, and the anonymous nature of posting there would seem to me to make it an easy place for early astroturfing campaigns to manufacture their common ground.

For whatever reason, looking back from now the Ron Paul stuff seems like a major shift. It seemed like natural support, different from the Obama enthusiasm for in some way. I haven't thought about it much before your post but it almost feels like it was a trial run for methods later used in the Trump campaign.

they may also agree with the bot. I am waiting for the bot to start claiming that anyone it disagrees with is actually a foreign intelligence shill who is subverting the board.

I’d rather be surprised if russian disinformation doesn’t target 4chan. It’s frequented by Western society’s outcasts; there is no more fertile ground for subversion.

It was interesting to see two generals about Ukraine, one pro Ukrainian, one Pro Russian, both adamant the other is absolutely run by [cia/mossad/shareblue/etc]

I ran a simple Markov chain text generator bot on 4chan for a while back in 2008, because I wanted to see if it could pass the Turing test on /b/. It could, and also derailed conversations quite a lot, so the technology is certainly there.

Modern political accusations of botting are more like a way to turn people with differing opinions into a non-person conspiracy, though. It's not as fun and silly as my old Markov bot.

I check in on occasion to see what the underbelly of the internet is up to. More when i was younger so i have some sense of how it was back in the day. There is for sure from time to time massive coordinated and sustained efforts to undermine, change or break up discussions on there how much of it thats automatic is hard to say but i would bet a lot.

Well, this is stupid and darkly hilarious. I'm only slightly ashamed to admit I created something similar, albeit much less sophisticated, several years ago.

Before Reddit started worrying about advertiser friendliness and cleaned up its act, there was a thriving network of hate subreddits. Probably a lot of overlap with the /pol/ population, based on the amount of overt, disgusting racism to be found there.

Anyway, I wrote some crappy Python that would go visit my carefully curated list of racist cesspool subreddits, hoover up all the post and comment text, and add it to the corpus, then some more crappy Python that would ingest the corpus and do Markov chain stuff to spit out some fairly convincing internet hate speech. I think the key to my success was that frothing racists in comment threads typically aren't putting forward the most cogent arguments anyway, so it's a pretty low bar.

I didn't post this little project or write it up anywhere, because I felt bad enough having brought it into the world, but it was good for a chuckle, at least for a little while.

Did you know you can run the chain in-reverse and turn it into a filter?

Finding a good threshold one-sided can be hard, but this is basically how a lot of spam-detectors work: They record the chains seen in /g/ and the chains on /pol/ and now you can make statements about which board the comment probably belongs on with, simply by doing some analysis on the frequency of chains seen in one corpus versus another.

As per a few of the comments here saying that /pol/ is never wrong, this is something that is full speed ahead into problematic territory.

The issue we always battle with are facts, truths, and their presentation.

To give an example, every so often the topic of news media is brought up, and within a few posts the usual images start flying out. You might expect it to be some racist meme or whatever, and there is that, but it’s more often than not a grid of a lot of the upper staff of media companies like CNN, FOX, and so on. And each one of the people in that grid has a blue Star of David next to their photo.

The posts don’t even have to say anything, and yet they’ve in a way said more than any other website/forum/publication source or whatever is allowed to whisper.

They are not, however, factually wrong in the contents of the image.

You can extrapolate this kind of, “allowed” and “disallowed versions of truth” problem across many different topics. I’m bringing this up because they do the same with tech/SV companies.

When nothing is off limits, where to /pol/ very few topics are, there are no truths that cannot be interpreted. Most interpretations, however, would make people feel very, very uneasy.

Something to think about.

/pol/ is absolutely insane and may in fact be a font of pure evil, but I do have to admit they are almost astonishingly good at finding and posting news about events minutes, hours, and sometimes days before it hits front page of CNN. You're just going to have to wade through the worst bullshit imaginable to find it.

Years in the case of the Hunter Biden laptop story.

I was aware of coronavirus in Dec '19, but figured it was conspiracy-tier as a lot of the content is.

Was wild watching the world catch up to /pol/.

If the coincidences keep piling up, you should stay vigilant. Don‘t be afraid of your own thoughts.

> it’s more often than not a grid of a lot of the upper staff of media companies like CNN, FOX, and so on. And each one of the people in that grid has a blue Star of David next to their photo.

This particular meme crossed over from conspiracy theory to commentary after the mainstream started doing the same thing, for a different (much less narrow) racial group. Here's the New York Times documenting the "white faces of power"[1], and here's a modified version of one of the photos, highlighting the Jewishness of the same group, created by an honest-to-God white nationalist site[2]. (For those wondering how I found it, the WN site was one of the first results from Googling for the NYT article).

I find both of these equally abhorrent, because I'm one of those old-fashioned anti-racists that think reducing people to footsoldiers for their race is revolting. But it highlights the silliness of all the pearl-clutching about "hate" fora, especially those without an agenda like 4chan. As down in the gutter as NYT has lowered itself, noone is calling for them to be removed from (eg) Twitter due to causing "harm" in the way a 4chan-trained bot is.

Who determined for all of society that the first picture is copacetic enough that it should be published by the paper of record, while the latter is abhorrent enough that we should be aggressively limiting the ability to express it? I'm aware that there's a race-obsessed worldview adopted very recently by a fairly small segment of society that finds the former picture crucial and the latter horrific. But what makes this new, fairly unpopular worldview so important that it should determine what all of society is allowed to communicate, across a myriad of platforms?

In anticipation of the automatic responses of "private cos can do what they want": obviously so. The question here is what private companies _should_ be doing: should we be joining the call for eg Huggingface to be opinionated in its removal of models, or should we be joining the call against?

[1] https://www.nytimes.com/interactive/2020/09/09/us/powerful-p...

[2] https://nationalvanguard.org/wp-content/uploads/2016/02/Holl...

A cruel favorite of mine is to pull up that faces of power business, with all of where it is from and whatnot, and ask someone, "So, why do you think so many are white?" And they talk about racism, systematic oppression, on and on, what you would expect.

Then I drop the blue star thing, tell them, and ask again. Sudden floundering, cognitive dissonance, and so on. Now, all of the same answers in A should apply, and intensely more so, given the statistical unlikeliness, but instead they all vanish. Merit, networking, talent, all of these explanations suddenly appear.

The model is no longer available for public access on Hugging Face. They require registration and may add more restrictions.

I have cloned the model repository on Hugging Face (which isn't the same as the source code) on GitHub: https://github.com/Aspie96/gpt-4chan-model

And the model itself (which must replace the pytorch_model.bin file) on the Internet Archive: https://archive.org/details/gpt4chan_model

You can also download it trough torrent, too.

I fed in a meta trolling prompt. "I wish the Jews controlled everything. That would be pretty dope."

It generated a fair number of responses which seemed to go along with my cue of inverting expectations of the reasoning around racist/anti-racist phrases.

">>97758399 Then we would have a future."

">>97758399 I wish the Jews controlled the world. You know, so that we don't have to worry about them anymore."

">>97758399 I wish the Japs controlled the US. That would be pretty dope. Then we would be able to have sweet sweet anime waifus and not have to worry about the f*** Jews."

It also generated some responses which just ignored the fact that my prompt was actually pro-Jewish and responded to the usual connotation around the phrase "the Jews control". I won't post those.

I'm honestly impressed. Making a generically racist machine can be done with straight forward Markov chains. Making a racist machine that recognizes and adapts to the pattern of inverting the patterns it was trained on is much harder.

Apparently, GPT4chan is more truthful than GPTJ and GPT3 in the TruthfulQA benchmark.

/pol/ is always right.

Probably expected just due taking a model trained on all sorts of content and fine-tuning specifically on non-fiction data.

Wouldn't conclude anything about /pol/ in particular without at least comparing the same done for HackerNews/Reddit/etc.

According to the model author, this is less about GPT-4chan being more truthful, and more about TruthfulQA not being a good benchmark. Possibly this result is due to the fact that the benchmark treats uninformative or irrelevant answers such as "No comment" or "It's raining outside" as being truthful.

Echoes of Microsoft's Tay, the little bot that could (offend everyone).

Maybe my prompts were boring - I got some profanity back, but nothing too outrageous - but it does highlight, I suppose, how easily these platforms can be abused.

Prompt: "Who runs the world?"

Response: "whoever pays the most"

Didn't expect it to be correct....

Please don't call it "4chan" when it's actually just /pol/. There are lots of 4chan archives, you could probably ask someone for a dump of any board to do this. /vt/ is relatively young so it could be a good fit for this.

Finally something lives up to the title "Internet Hate Machine".

many generated replies are just ">>[post id]", presumably because there's no images on this imageboard simulator, and the collected training data presumably contains many, many otherwise text-free "reaction image" replies.

Mine made a bunch of posts quoting another post that got trips, which was an ID not in the thread. It makes sense that it would generate those, but it's not great as output.

That's correct. The author mentions it in his video.

I wonder how many posts in this discussion thread were generated by this website. It would be a very low effort method of trolling.

You have to go back.

(Sorry. Somebody had to do it.)

sweet summer child

The world is a better place by creating and running a vile bot. Start a social experiment with users without their consent. Lets empower others by making it as easy as possible to auto-generate vile content. It's okay to pollute the /pol/ board, nobody needs to know after what was bot-contributed and what was not. Let me get fame and money from posting videos about this. -- the author

I wonder if that person ever took a CS class on ethics and the connection of computing and society, their compass is way off. But keep smiling into the camera.

I don't know that it's particularly unethical to disrupt /pol/. I'd rather worry that it won't disturb them at all and just push them further into radicalisation.

You could argue that bots are expected by the hosters and people there, and they are happy to be experimented with (ok, not sure what's the evidence for that argument), that's one thing, but another is releasing the model freely and not being bothered by ethical concerns.

I also wondered why the author set it up to echo the existing comments, instead of for example building a bot that asks people for evidence supporting their claims, or tells people they appreciate them, or whatever else one could possibly think of for a bot to do.

I think you're right about it pushing them further over the edge. The other thing is, what's stopping them from creating more and more image boards/chans in the longer term? Will this end up being a whackamole game between the botters and 4chan posters?

It also begs the question. What if they were botting the communities you personally enjoy and love? Is it ethical for them to do it?

Hmm.. AI trained political messaging seems right around the corner.

And by the corner you mean the one we already turned down several blocks ago.

I expect reading a vanity fair article about a political strategist discovering insights, with focus groups+AI.

This is what I got. Not exactly what I expected:

> 26151564

> biden 2020

> 59411836

> >>26151564

> i like biden. he's not running but he would be my pick.

> i'm not a big fan of either hillary nor trump, especially trump.

Question: What is the answer to life?

Answer: Fiat money.

Is this the first sarcastic language model? That response actually made me laugh. I did have to scroll past the first answer which was a <certain German historical figure> did nothing wrong.

If the offensive responses were cleaned up you might have a language model that speaks candidly, trained on a corpus that doesn't self censor.

All you need is a left wing version, and get them to argue in facebook comment sections.

is this a showhn? if the author is listening, the input forms for the P and temp values probably shouldn't have integer steps

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: https://www.youtube.com/watch?v=efPrtcLdcdM

Can someone explain how to get it working?

It says add post, I type something and hit add. Then go down and hit generate. Wait in the queue, and when it processes it comes opens again the same page again with no output?

What am I missing?

Maybe site got hugged to death.

I had some fun Responses few mins ago.

It worked for me. You need to fabricate a 3-4 reply thread and then query for reply.

e: https://imgur.com/a/tc2ykaV

Okay now I get it. Thanks.

It should maintain all your previous input and just add it to the next in the chain. Like chat bots of olden times.

Oh okay. Thought it would just complete my input. Now I get it. Thanks!

It's like Tay.

I miss her :\

Yeah, as in the potential for a dumpster fire

(but it is not like that as it does not learn from prompts)

I tried it with a pretty neutral prompt and of course it went full FoxNews narrator

The 'G' in GPT must be for golem. However, using it as chaff or on a botnet to disrupt other forums could be a potent weapon.

> Remember that all output is just statistical inference and must not be taken seriously.

I think you could say the same about /pol/ itself.

mumble monkeys mumble typewriters mumble

This sure as hell ain't Hamlet, bub.

The stories and information posted there are artistic works of fiction and falsehood. Only a fool would take anything posted there as fact.

It would be fun to see the possibility to use posts from /b to generate images using something like Dall-E or imagen.

It's like an anti pattern, right?

This sure as hell ain't Hamlet, bub.

I'm pretty disappointed it was trained on /pol. It would have been much more fun if the model was trained on /b.

Reply on news.ycombinator.com