Grok has seemed unusually prone to bizarre behavior, even if you don't count the times that its system prompt was modified (like the infamous "please talk about

Grok has seemed unusually prone to bizarre behavior, even if you don't count the times that its system prompt was modified (like the infamous "please talk about white genocide" incident in May). Part of the problem seems to be that it's being trained on recent Twitter posts, resulting in a weird sort of feedback loop when users discuss Grok's behavior.