X's Grok AI is suddenly hyper-fixated on South African farmers | Hacker News Reader

X's Grok AI is suddenly hyper-fixated on South African farmers

bsky.app

43 pointsjsheard21 comments

Some more examples of this happening:

https://xcancel.com/grok/status/1922668715533083121

https://xcancel.com/grok/status/1922651218595439063

This is an example of why you shouldn't link to X or anything that looks like X.

Here you've got some guy who is talking to himself while pretending he's talking to others, not giving enough frame to understand what really happened. It looks like he showed it a picture and Grok started blabbing about "farm attacks" but what do I know?

Move on folks, nothing more to see here.

I linked via xcancel to get around the login wall, and it seems clear enough to me? The OP posted a photo, someone else pinged Grok to ask where it is, and Grok went off on a tangent about SA farm attacks apropos of absolutely nothing.

IMO it's notable because it looks an awful lot like Elon Musk rigging xAI to aggressively push his own talking points, ineptly for now, but it'll only get harder to spot.

It still doesn't seem very interesting, even assuming the theory is right. Like, what's the interesting conversation you envision would happen here?

It's not an interesting conversation. It's a live example of what's happening with grok right now, and it's not hard to parse:

    1. Person A posts something
    2. Person B calls @grok to weigh in
    3. @grok starts babbling about South Africans

It probably is representative of training data based on Musk’s bent world view but doesn’t come across as effective propaganda. If it said some particular farmer got attacked at a particular time and it happens at a certain rate or so and so threatened more attacks or something it might be persuasive. As it is it just looks like an unhinged part of the cringiverse.

It could also be based on what Grok is being fed from the users... IIRC, there was an early AI bot from Microsoft intended to have the mentality of a young teen, and within a couple days, the bot was a racist slut just from interactions.

It depends on how the bot takes in the information it is interacting with, not necessarily the directions of a single person...

I mean, someone asks, "What's going on with the farmer's in South Africa?" and the bot has seen all the unhinged posts about the topic, and injects that information into the position.

It's not necessarily malice, but yet another example of why I generally don't trust AI platforms and would reject actions based on AI results without external validation.

I'm not having trouble parsing it, I'm asking what would make this an interesting HN post.

It looks like the system prompt may have been modified, or that it may be some kind of attack on the system prompt, or possibly someone attacking the training data, or it may be totally organic somehow, in which case it's still interesting how topics can leak into responses to unrelated questions.

These are all interesting topics for the crowd on HN to weigh in on.

If I searched for some arbitrary thing on Google right now and the AI generated results gave me a bunch of unrelated slop about South Africa, we would certainly be discussing how that might have happened. This is no different.

'Musk fiddles with twitter to match his weird views' is a very repetitive story that's been on HN a bunch, including recently so I guess I don't see the interestingness in this case.

The 'attack' thing is your gloss, I don't see it appearing in any of the reporting.

A story that I think is clearly worth discussing on HN is "what seems to have happened with grok here?". Musk fiddling with the system prompt is only one of the possibilities. People here can make arguments for or against that theory or others.

But also, it's news worth discussing on a technology news discussion site if the owner of one of the largest technology platforms by reach is manipulating the information on that platform in a new way, even if there have been lots of other discussions about them manipulating things in other ways.

Notably, I think the only uninteresting threads here are the ones like this one that are questioning why this is news. Yawn.

I'm guessing that your view aligns with the theory that novelty is one of the more important measures of a HN post. In other words, a topic that "is a very repetitive story that's been on HN a bunch" is not novel, so is less valuable on HN. In any case, whether that is your view or not (sorry if I misinterpreted you), I think this is too narrow a view. A topic can be one that has been on HN a lot, but still be novel, interesting, fresh if the new story has some different twist or nuance to it. Also, not everyone reads HN to the same degree, so what is "very repetitive" to one set of readers may not be to another. I know that some users flag certain posts (and mods do not subsequently unflag them, even on appeal) because of their perception of a post's lack of novelty, but this seems overly censorial when these users could skip reading it, not upvote it, etc.

> I'm not having trouble parsing it

It seems like you might be. Several people have shared their opinions on this topic, some directly to you, and in your other posts you insist that this shouldn’t be on this website because you, user pvg, does not want to see it.

The issue to parse here is that this wasn’t emailed to you, it is a discussion on a public forum that you clicked on and posted in several times. I would also be very confused if I thought I was being forced to read things on this website against my will.

It doesn't matter whether it is Mastodon or the "archive.today of X", almost anything contentious on those platforms turns into an infinite regress of finger pointing.

I think the first I heard of "gamergate" was that some guy, maybe in Maine, seemed to think that a bunch of radical feminists flipped over his car but it was an incredible story told from the viewpoint of somebody who already knows the moral status (always double plus good or double plus ungood) of 1000s of people.

People who are addicted to those platforms seem completely oblivious to it but the user interface on all of them for threaded discussions is terrible.

>think that a bunch of radical feminists flipped over his car

Like the cartoon character She-Hulk? That's a funny image.

My image is that either (1) it was a lot of people or (2) he did it himself or (3) was suffering from delusions.

Was the guy who flipped his car the comedian?

Yes, pay no attention to the man behind the curtain, nothing to see here! /s

> Tweet not found

Were the originals deleted?

Yep, it looks like it. Some threads pointing that out are here:

https://x.com/devahaz/status/1922765156699767286

https://x.com/colin_fraser/status/1922761230269251785

These two slightly newer, but flagged, HN posts have larger comment threads about this:

https://news.ycombinator.com/item?id=43993332

https://news.ycombinator.com/item?id=43987266

Reply on news.ycombinator.com