Block a user
Reinforcement Learning with Human Feedback
Here's an idea: we already have a ton of training data, with reactions and all. If we presume more reactions are given to "better" messages, whether that means provocative, funny, or something…