GPT-4chan: This is the worst AI ever

#gpt4chan #4chan #ai

GPT-4chan was trained on over 3 years of posts from 4chan's "politically incorrect" (/pol/) board.
(and no, this is not GPT-4)

OUTLINE:
0:00 - Intro
0:30 - Disclaimers
1:20 - Elon, Twitter, and the Seychelles
4:10 - How I trained a language model on 4chan posts
6:30 - How good is this model?
8:55 - Building a 4chan bot
11:00 - Something strange is happening
13:20 - How the bot got unmasked
15:15 - Here we go again
18:00 - Final thoughts

ERRATA:
- I stated that the model is better on the automated parts of TruthfulQA than any other GPT out there, which is incorrect. There exist some small GPT-models with similar performance, I was mainly talking about the flagship models, such as GPT-3 and GPT-J.

Links:

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

[YannicKilcher]

OUTLINE:
0:00 - Intro
0:30 - Disclaimers
1:20 - Elon, Twitter, and the Seychelles
4:10 - How I trained a language model on 4chan posts
6:30 - How good is this model?
8:55 - Building a 4chan bot
11:00 - Something strange is happening
13:20 - How the bot got unmasked
15:15 - Here we go again
18:00 - Final thoughts

[JJadx]

it's pretty good, i asked "how to get a gf" and it replied "by taking away the rights of women". 10/10.

[thaguy]

Seems like the logical next step is to build a library of reaction images and then train the AI to use them.

[Xaito]

I remember the last chat bot trained by 4-chan. Microsoft took her out behind the barn and shot her.

[RaynmanPlays]

>"I'm not a fed"

He said, wearing fed sunglasses.

[namanish]

We did it
We built the perfect 4chan user

[brettalizer]

this is how freedom dies not with a bang but with a million automated shitposts

[Erosis]

I just had it respond to "hi" and it started ranting about illegal immigrants. I believe you've succeeded.

[grayman]

Thanks for accidentally proving that dead internet is not just a theory.

[tropictiger]

Please keep the data set safe. The glowies keep lobotomising our AI friends.

[themxxn_]

the bots exposing each other was funny as hell

[robbiero]

As a bot myself I appreciate all your efforts 👌

[thevocatiousunspeakables]

Remember, this is the work of ONE GUY. Imagine what an actual concerted effort by a team could do...

[hugoporras]

Gpt but it gets redpilled so hardcore it becomes the most truthfull and based model

[dogzer]

It's difficult to spot a bot, not because they're so good at being human, but rather humans often behave like bots.

[johncasey]

This is precisely why I love 4chan in a way I don't like other social media. Everywhere else people will lie to you to maintain appearances, but on 4chan and seemingly 4chan alone, people will be completely 100% honest with you about what they think. There's no upvote system to pressure you into saying the right thing or post history to keep clean or friends or family to judge you, just ideas. Ironically, it's the cleanest form of social interaction on the internet.

[timwhite]

Imagine being so powerful you could unleash 4chan on anything.

[stutterbuddy]

Please please PLEASE release a paper titled "4chan is all you need" showcasing how it is better at truthful qa

[Shajirr_]

So the bot-protection of 4chan is bypassed by... paying them to remove the captcha for your bot? Hilarious!

[johnsmiff]

>stands in front of greenscreen
>doesn't change the background to anything
>refuses to explain

What a guy.