very weird thing happening with Grok on Twitter/X lol

Elon Musk's AI chatbot can't stop talking about South Africa and is replying to completely unrelated tweets on there about "white genocide" and "kill the boer"

@MattBinder Uh hmm that's a strangely specific topic for it to be taking an unprompted interest in
@liquor_american @MattBinder That's the kind of things that happen when you put too much of a single topic in the system prompt, especially when it contradicts the training data. aka AIs misbehave when asked to go against their training data

@aris @MattBinder I'd argue that LLMs misbehave consistently under all conditions, but we've decided it's an acceptable level of misbehavior that can be tolerated.

Nope nope nope.

@liquor_american @MattBinder Depends what the level of expectation and trust you have on its answer. Provide a slightly different wording for your party invitation or recipe of pancakes? who cares.
Fact-checking (mis)information? I agree 100%.
@MattBinder @GossiTheDog sounds like Grok is giving the first part of the reply and then the second part is hardcoded?
@MattBinder Elon testing Grok in prod, and this being responses to leftover prompts? Maybe?
@MattBinder it looks like someone messed up hurriedly the system prompt with a very specific "both narratives' instruction

@MattBinder

Most likely the system prompt now contains instructions to handle that topic in a special way ('Both sides...'). But the problem is that the topic is in the prompt all the time then, and will crop up without user prompting.

@mhhwhitney @MattBinder Yep, think that's pretty likely. They shoved some special instructions into it and didn't do a good job.

(Edit: Quite apart from the fact that the thing it's saying here is carefully worded to give the illusion of "both sides"-ing this while actually coming down hard on one side.

@MattBinder This broken man is dragging the entire goddamned country down his rabbit hole.
#USpol
@MattBinder South Africa is directly in the system prompt I guarantee it. Meaning Elon and top Twitter brass are rather blatantly putting their thumb on the scale of public discourse and barely even trying to hide it. What a dope.

@MattBinder This is a very good example of what can happens when you overrepresent a subject in the fine-tuning data.

So my two cents, if this is correct, is exactly that.

@MattBinder turns out I gave them too much benefit of the doubt. It looks like they actually did the thing every expert on the subject would know is a bad idea. And added it to the system prompt 😮