Exploring GPT-4chan: The Most Racist Language Model Ever Conceived?

GPT-4chan is Yannic Kilcher's language model trained on 3.5 years of /pol/ posts on 4chan using machine learning. It has a tendency to be very racist, but this is not always the case. For example, by avoiding controversial topics such as race, sex, gender, politics, religion, economics, the police, the environment, abortion, mental health, and drugs, you can have a nice, polite conversation.

Empty replies are because the bot has learned that 4chan users will sometimes post images unaccompanied by text. However, the bot is unable to post images, so it just quotes the user with nothing.

The default settings are very effective. I suggest you play around with them and see what happens. Here is how each affected my experience:

Length: Shorter is quicker to load but with simpler generation. Longer is slower but with more complex and realistic posts and replies.
topP: Lower causes the replies to copy the initial post or each other. Higher leads to more unique posts and replies.
typicalP: Similar to topP, but less impactful.
Temperature: Lower temperature is less random in replies. Higher temperature is more random.

UPDATE: As of 6/22/2022, access to this model has been indefinitely disabled by Hugging Face. However, there are backups and torrents available to use at your own risk.

Discussion w/ Backup: https://huggingface.co/ykilcher/gpt-4...

Try It Yourself: https://gpt-4chan.com/
Official Video:    • GPT-4chan: This is the worst AI ever  
Documentation: https://huggingface.co/ykilcher/gpt-4...

"Neolith" by Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 3.0 License
http://creativecommons.org/licenses/b...