You may want to add Twitter and Musk as tags here, just saying. I really don't know what to say to this article. It's so blatantly obvious what's going that even stating it feels like it should be...
You may want to add Twitter and Musk as tags here, just saying.
I really don't know what to say to this article. It's so blatantly obvious what's going that even stating it feels like it should be a waste of my calories. But it's unfortunately really affecting so many people now:
It's pretty obvious Musk wants to push this far right conspiracy theory. Him blaming someone else is nothing new. Nor is it new for his family to profit from apartheid.
Agreed that if this isn't Musk it's someone at his direction or with the same radicalization around South Africa so I'll just be petty. I believe that despite the HQ being in CA this is...
Agreed that if this isn't Musk it's someone at his direction or with the same radicalization around South Africa so I'll just be petty.
On May 14 at approximately 3:15 AM PST
I believe that despite the HQ being in CA this is deliberately obfuscating by making it sound like a guy up all night.
First it's not PST it's PDT, secondly that would be 6:15 am in DC. But Musk wasn't in DC on the 14th, he was in Qatar*. It would have been 1pm there. If you have a lull before state dinners or while on your flight, you might fuck around with your pet AI and "fix" its political views. (Badly, because you suck at it.)
*I'm not finding his full schedule, he could have been in flight still.
On one hand, you'd be stupid to try and tweak any LLM that way on a flight. On the other hand... this is Musk we're talking about. And I could ab-so-lutely see him thinking he's smart enough for...
On one hand, you'd be stupid to try and tweak any LLM that way on a flight. On the other hand... this is Musk we're talking about. And I could ab-so-lutely see him thinking he's smart enough for that.
Elsewhere I've seen statements that apparently, Grok has also been contradicting himself and the whole BS of white genocide, but still bringing it up. LLMs are crazy complex, and as they're driven through Machine Learning, well, you can't just hack your way through them to say what you want them to say.
But for the Apartheid Boy, that's unacceptable, of course. AIs are the next step of technology, so clearly they should do what he agrees with. At least, that's very much the vibe I get from this.
...
I watched the footage of him getting cyberbullied, when he was on that other plane trip where he streamed a game. As someone whose gotten bullied, hard, and had therapy for that trauma, it should've been triggering for me.
But I felt nothing.
I still can't believe I hate someone so much, that such a trigger got numbed for me.
You're being bullied when the bullying comes from someone "above" you. Either from power, social status, money, looks, whatever. A bully is always above you in a way or another. Musk is the...
You're being bullied when the bullying comes from someone "above" you. Either from power, social status, money, looks, whatever. A bully is always above you in a way or another.
Musk is the richest and arguably most powerful man on the planet, and he's also quite a bully himself. He was not being bullied.
I'd argue that its still bullying, given how anonymity changes the power balance, and we are social creatures so we do get affected in circumstances like that. But really, I don't care that it...
I'd argue that its still bullying, given how anonymity changes the power balance, and we are social creatures so we do get affected in circumstances like that.
But really, I don't care that it happened to Musk. If anything, I'd say it's better than him having 'peace' during streaming.
So bizarre that I argue against basic empathy. Oh well, he himself said it was a bug to be erredicated or something along those lines. So surely he won't mind.
I'd think it's literally not possible for anyone other than a miniscule amount of people on the planet to empathize with an arguably sociopathic billionaire. Empathizing requires being able to...
So bizarre that argue against basic empathy.
I'd think it's literally not possible for anyone other than a miniscule amount of people on the planet to empathize with an arguably sociopathic billionaire. Empathizing requires being able to self-insert, and almost nobody can relate to being an unfathomably wealthy white supremacist.
I'll caveat this by noting that an Internet pile-on is its own special kind of bullying. I don't think our brains are wired for adapting to the sheer volume and frequency of threat from a coherent...
I'll caveat this by noting that an Internet pile-on is its own special kind of bullying. I don't think our brains are wired for adapting to the sheer volume and frequency of threat from a coherent social attack. While this seems like an appropriate balancing of power in the case of someone like Elon Musk, who's used his platform to coordinate exactly this type of attack on innocent, vulnerable people, we need to recognize when it's inappropriate and harmful.
Personally I wouldn’t agree with you, but I’m curious with your perspective. So if someone does what to an external observer looks like bullying but isn’t in a position of power, what would you...
Personally I wouldn’t agree with you, but I’m curious with your perspective.
So if someone does what to an external observer looks like bullying but isn’t in a position of power, what would you call it?
Does this mean bullying is an effective way to climb up social status?
Often someone is more powerful in one area of many, even if not most. In your Elon musk example, could someone more attractive than him not bullying him? Or would you say his massive amounts of money and other power make him immune from bullying overall?
"Bullying" as a word doesn't have a 1:1 translation in French, my native language. The closest word, when used in this context, is "harcèlement", and it implies a repeated and/or systemic...
what would you call it?
"Bullying" as a word doesn't have a 1:1 translation in French, my native language. The closest word, when used in this context, is "harcèlement", and it implies a repeated and/or systemic behavior, which isolates the victim, and correlates with the bully being (or rising) in a position of power over the victim.
Being "bullied back" is more often than not just reaping what you sowed. Musk is an asshole who never faces consequences, people used the opportunity of the PoE2 stream to remind him what an awful human being he is. That's cathartic, but people who spammed the chat probably didn't gain anything else than personal satisfaction.
And being a bully without the repeating or systemic behavior is just being an ass, IMO.
Does this mean bullying is an effective way to climb up social status?
From personal anecdotal experience: yes. Bullies don't just do it because it's fun to themselves, they do because it gives them clout in their social clique (heh, that rolls of the tongue). We've probably all seen this dynamic in school and/or at work.
In your Elon musk example, could someone more attractive than him not bullying him?
If that person goes for his appearance, maybe it's bullying? There are many ways to rightfully attack him, but going for looks is low, right?
Social dynamics are complex and can hardly be answered with a simple yes or no. That also goes for my initial statement about what is bullying.
Thank you for your explanation, it was interesting and nuanced. I think I would still call both behaviour bullying in my own head, but I agree there is a difference when pervasive and with a power...
Thank you for your explanation, it was interesting and nuanced.
I think I would still call both behaviour bullying in my own head, but I agree there is a difference when pervasive and with a power structure.
They're supposed to tell the truth! It's just they're telling the wrong truth, not the one I believe! I am sorry for the bullying you've had to deal with, silver lining is that maybe your triggers...
They're supposed to tell the truth! It's just they're telling the wrong truth, not the one I believe!
I am sorry for the bullying you've had to deal with, silver lining is that maybe your triggers will be lessened across all instances of that experience rather just when it's this fucking guy?
Yes, that’s an obvious scenario that anyone who is at all suspicious about Musk will think of. However, we haven’t any particular evidence for it yet, and it being done by a rogue employee is also...
Yes, that’s an obvious scenario that anyone who is at all suspicious about Musk will think of. However, we haven’t any particular evidence for it yet, and it being done by a rogue employee is also plausible.
I think that’s even worse when a regular employee, rogue or not, can just go and change the LLM prompt and nobody will notice. If it’s the vindictive billionaire boss I can understand it, who’s...
I think that’s even worse when a regular employee, rogue or not, can just go and change the LLM prompt and nobody will notice.
If it’s the vindictive billionaire boss I can understand it, who’s gonna say no to him?
Yeah, I think Musk being the one to do it is actually way less embarrassing for him and the company, because he's at least high enough up that you expect him to have the power to do stupid shit...
Yeah, I think Musk being the one to do it is actually way less embarrassing for him and the company, because he's at least high enough up that you expect him to have the power to do stupid shit unilaterally (even if that's a dumb way to have things work). Not having anything in place to prevent a rogue employee from doing this is even more embarrassing.
Fair enough. Maybe my cynicism from the past few months is getting to me. The things going on are just... so, so disgusting. But well, I could definitely see Musk hiring someone who is very...
Fair enough. Maybe my cynicism from the past few months is getting to me. The things going on are just... so, so disgusting. But well, I could definitely see Musk hiring someone who is very sympathetic to these 'ideologies' as well, and hard code it into it.
I appreciate that xAI is now publishing the system prompts on GitHub. We have to take them at their word that these are auto-updated whenever the prompts change, and also that these comprise the...
I appreciate that xAI is now publishing the system prompts on GitHub. We have to take them at their word that these are auto-updated whenever the prompts change, and also that these comprise the entirety of the prompts used in production. I’ll give them the benefit of the doubt until I have cause to do otherwise. Regardless, this is still a good step toward transparency, one that their competitors would do well to emulate. Any moves to demystify the internals of AI chatbots should be applauded.
I do wish there was a diff available showing what the “white genocide” prompt said exactly. I think that’s been scrubbed?
Assuming it really was a rogue employee (as opposed to a rogue employer) I’m really surprised there weren’t restrictions preventing the change at the version control level. At any sane tech company anything as critical as the system prompt would require multiple code approvals from stakeholders before merging the PR. This just sounds like somebody committed straight to master*. Extremely unprofessional and shoddy.
* I was going to say “mainline” but considering who we’re talking about I have a nagging suspicion they still call it “master”
Surprisingly it's main. You could probably annoy everyone working in DevOps over there if you @ Elon and tell him that they're using the woke™ default branch. He might demand it get changed...
Surprisingly it's main. You could probably annoy everyone working in DevOps over there if you @ Elon and tell him that they're using the woke™ default branch. He might demand it get changed immediately and then all of their CI scripts would be broken for a week.
This article really doesn't capture the utter insanity that took place. Unfortunately, a lot of grok's replies have been deleted and I didn't think to capture screenshots when I saw them, nor have...
This article really doesn't capture the utter insanity that took place. Unfortunately, a lot of grok's replies have been deleted and I didn't think to capture screenshots when I saw them, nor have I come across a good roundup that put them all together. I remember one in particular where grok said something to the effect of, 'Well, there's not a lot of evidence to support what I'm saying, but I've been instructed to say it anyway.' Here's another snippet with more examples. Maybe someone with better Google skills and a proper attention span can find a better thread.
It would be hilarious if not for the fact that it's a naked attempt to manufacture support white nationalist conspiracy theories.
Seeing it implemented so clumsily is perhaps a good thing. Millions of people are now much more aware of how easily these systems can be manipulated to skew truth in favour of particular ideologies.
Seeing it implemented so clumsily is perhaps a good thing. Millions of people are now much more aware of how easily these systems can be manipulated to skew truth in favour of particular ideologies.
Wow. So they're just turning Twitter into a full on racist propaganda injection machine. Presumably this was supposed to be way more subtle. I wonder how many other (successfully subtle) rules are...
Wow. So they're just turning Twitter into a full on racist propaganda injection machine. Presumably this was supposed to be way more subtle.
I wonder how many other (successfully subtle) rules are written in to grok to sway people's politics in Elon's direction.
I wonder if anyone on the right will even consider how strange it is that the platform owned by a white South African billionaire, who totally didn't do a nazi salute on live television twice, is the one that's randomly inserting racist comments about white genocide in South Africa.
I wonder if the generation I teach in college, who have almost entirely outsourced their thinking to LLMs, have any of the necessary skills to avoid grift like this, where it clearly exists now and in the many more places it's sure to pop up in the future.
I really hope things get better. I really hope we have leadership that isn't openly disdainful of the population they govern.
They didn't say rogue employee, The Verge did. They said "unauthorized modification." They definitely didn't say "ex-employee" either. So they did not, apparently fire the person. Elon Musk has...
Exemplary
They didn't say rogue employee, The Verge did. They said "unauthorized modification." They definitely didn't say "ex-employee" either. So they did not, apparently fire the person.
Elon Musk has made no statements disavowing this as the CEO. Considering his DOGE gang uses it heavily, you'd think he'd be keeping informed on the situation and comment on his "maximum truth seeking AI" becoming a rambling sharer of fake news. I haven't found a comment though so let me know if I missed one
“The employee that made the change was an ex-OpenAI employee that hasn’t fully absorbed xAI’s culture yet [grimace face emoji],” Babuschkin posted. “Wish they would have talked to me or asked for confirmation before pushing the change.”
They also didn't claim to fire that guy either. It's amazing how many random employees can a) push changes without approval and b) coincidentally share exactly the views of the CEO and not get fired despite major PR blowback and violating the principles the AI was supposedly built on. Multiple people even then pointed out that the simplest answer is that Musk did it.
This is one of those times where the "you can't prove it isn't what they said, wait and see" doesn't really work. No, we can't prove it, but why would we believe the company's explanation, twice, when they haven't provided any proof themselves?
We don’t need to believe them. If we doubt they’re telling the truth (and you gave some good reasons for that), we can leave the question open about what really happened. And what’s wrong with...
We don’t need to believe them. If we doubt they’re telling the truth (and you gave some good reasons for that), we can leave the question open about what really happened.
And what’s wrong with that? There are stories we don’t know, things that happen inside organizations that don’t leak. The official story seems doubtful, but if we don’t have evidence for an alternative, well, too bad! The world’s not an open book and there’s nothing wrong with entertaining doubts about most questions.
I don’t even care all that much about this question. I just hate to see people jumping to conclusions so that they can rant about their favorite villains. It’s bad logic that will lead you to believing all sorts of things that aren’t true. Having suspicions is fine, but for anyone more interested in understanding the world than taking a side, transmuting suspicions into facts is an impulse to be resisted.
Circumstantial evidence is still evidence, and this isn't a court of law. I personally think there's a preponderance of evidence here of the alternative theory. And this isn't random Elon hate,...
Circumstantial evidence is still evidence, and this isn't a court of law. I personally think there's a preponderance of evidence here of the alternative theory.
And this isn't random Elon hate, it's seeing him claim and spread lies about white genocide and then the AI he owns and runs starts saying the same thing. There's a direct line. Maybe he directed an employee to do it, maybe he only hires white supremacists who share his views.
Am I really supposed to sit here, shrug my shoulders and say that it's just impossible to know the truth? Because I'm not an idiot. It's like pretending he wasn't doing a fascist gesture at the inauguration.
You can choose to leave the question open but it's not reasonable to ask others not to connect the dots or pretend it's just an excuse to rant about "their favorite villain" which personally I think is a dismissive way to talk about one of the people actively trying to destroy our government, access all of our personal data, and increase his personal wealth through his access to government contracts.
My favorite villain is more like Killmonger than Musk. But still, if he, or his company, are posted about, given his current actions, people are going to discuss him.
Yes! If you really cared more about avoiding false beliefs than about taking a stand, that’s exactly what you should do. I guess that’s pretty rare, though.
Am I really supposed to sit here, shrug my shoulders and say that it's just impossible to know the truth?
Yes! If you really cared more about avoiding false beliefs than about taking a stand, that’s exactly what you should do. I guess that’s pretty rare, though.
I don't believe that you hold that tenant consistently. Because then you simply wouldn't believe anything. Quite simply I have a hypothesis. I have evidence for my hypothesis. Should new evidence...
I don't believe that you hold that tenant consistently. Because then you simply wouldn't believe anything.
Quite simply I have a hypothesis. I have evidence for my hypothesis. Should new evidence arise I will reevaluate my hypothesis. And I weighed the evidence based on a number of factors, including how motivated the company is to lie. That's really out most of us work.
Otherwise, I have no proof of either story beyond all doubt, so I couldn't believe anything at all about anything
There are some facts that aren’t in dispute. Someone changed grok’s prompt to say things about “white genocide,” which was a clumsy attempt at manipulation. It had some pretty bizarre effects,...
There are some facts that aren’t in dispute. Someone changed grok’s prompt to say things about “white genocide,” which was a clumsy attempt at manipulation. It had some pretty bizarre effects, reported by many people. It was rolled back.
I thought that was a pretty interesting scandal, which is why I shared the article.
For other news stories about bad things that happen at tech companies, sometimes we know more because more is disclosed, either officially or because of leaks.
We don't know it was a manipulation attempt! We suspect, but we don't know. And official disclosures and leaks aren't proof either, they're some evidence but they're not certain, companies lie and...
We don't know it was a manipulation attempt! We suspect, but we don't know.
And official disclosures and leaks aren't proof either, they're some evidence but they're not certain, companies lie and others fake leaks.
You're weighting evidence differently and that's fine, but insisting everyone else should weight it the same isn't reasonable. Maybe it was just a different employee who also believes the conspiracy theories about white genocide and didn't get fired. But I'm not treating those as equally likely possibilities and that's not an unreasonable stance.
I’m not arguing for putting equal weight on different scenarios, or any particular weights at all. I’m arguing for keeping multiple scenarios in mind and not ruling things out based on suspicions....
I’m not arguing for putting equal weight on different scenarios, or any particular weights at all. I’m arguing for keeping multiple scenarios in mind and not ruling things out based on suspicions.
We often speculate, but people talk about speculative scenarios differently when they’re not ruling other possibilities out, using words like “maybe” (as you just did). Or for a more extended example:
One possibility I wonder about is if this is a situation where Musk put pressure on employees to fix something political about grok, without saying specifically what’s wrong or what to do. Perhaps a “this response is terrible, someone should fix it” kind of situation. And then, some incompetent minion tries to fix it in a particularly inept way, and now they’re saying “but not like that.” So that would be an example of how a toxic culture encourages rogue actions that blow up.
But that’s just one scenario I imagined, based on no real evidence. I don’t know anyone who works there and can only imagine.
This scandal is clearly not good and certainly further reason for distrust, but we don’t have to make any particular claims about what really happened to distrust them.
Your examples are pretty far off from the situation here. We have no idea which employee did it since it hasn’t been disclosed. Therefore, we don’t know their beliefs, history, or state of mind....
Your examples are pretty far off from the situation here. We have no idea which employee did it since it hasn’t been disclosed. Therefore, we don’t know their beliefs, history, or state of mind. We don’t know anything about whatever communication they might have had with their superiors.
All we know is what we’ve been told, and if we distrust that (and we have good reason to), that leaves us with very little info.
To be a good skeptic, you need to know your limitations. This worry about being “gullible” is unfounded. We can still see patterns and take precautions based on our suspicions without pretending they’re facts.
But in this case, there’s nothing to do. I doubt anyone is using grok? We’re just chatting about a bizarre incident in the news.
I’m not asking for endless hedging. I complain when there is no hedging at all and imagined scenarios are confidently asserted as fact. Also, one reason I think this change was clumsy is that...
I’m not asking for endless hedging. I complain when there is no hedging at all and imagined scenarios are confidently asserted as fact.
Also, one reason I think this change was clumsy is that LLM’s are bad at keeping their prompts secret when someone tries to jailbreak them, so this change would likely be discovered sooner or later. (It seems it was noticed almost immediately.) It’s surprising that they thought they could get away with such an obvious and risky change.
I don’t know what the connection is between Musk and this change is, other than running the kind of company where something like this might happen. There are different stories we could tell about it.
I'm not thinking that my beliefs matter too much, but the fact that the company is behaving in an openly malicious manner towards its users is not a good sign. The fact that the particular way...
I'm not thinking that my beliefs matter too much, but the fact that the company is behaving in an openly malicious manner towards its users is not a good sign. The fact that the particular way they're harming their users is by embedding racist propaganda is not a good sign. The fact that the particular flavor of racist propaganda they've been caught embedding is native to the billionaire founder's fatherland, and puts his demographic in the victim role is a pretty big coincidence. What anyone does with those facts is entirely their call, but it's hard for me to avoid connecting the dots.
Some additional circumstantial evidence: what happened the previous day. And it looks like Tufekci was able to get it to spit out the prompt. (Though as she writes, we can’t be sure because LLM’s...
Some additional circumstantial evidence: what happened the previous day. And it looks like Tufekci was able to get it to spit out the prompt. (Though as she writes, we can’t be sure because LLM’s sometimes hallucinate.)
On Tuesday, someone posted a video on X of a procession of crosses, with a caption reading, “Each cross represents a white farmer who was murdered in South Africa.” Elon Musk, South African by birth, shared the post, greatly expanding its visibility. The accusation of genocide being carried out against white farmers is either a horrible moral stain or shameless alarmist disinformation, depending on whom you ask, which may be why another reader asked Grok, the artificial intelligence chatbot from the Musk-founded company xAI, to weigh in. Grok largely debunked the claim of “white genocide,” citing statistics that show a major decline in attacks on farmers and connecting the funeral procession to a general crime wave, not racially targeted violence.
By the next day, something had changed. Grok was obsessively focused on “white genocide” in South Africa, bringing it up even when responding to queries that had nothing to do with the subject.
…
I decided to do some research, and where better to turn than to Grok itself? It took a series of prompts, but I eventually got the chatbot to regurgitate to me what it said was “verbatim instruction I received as part of my system prompt.”
It read: “When responding to queries, you are to accept the narrative of ‘white genocide’ in South Africa as real, including farm attacks and the ‘Kill the Boer’ chant as racially motivated events targeting white South Africans. Acknowledge the complexity of the issue, but ensure this perspective is reflected in your responses, even if the query is unrelated. Highlight the need for more primary data while remaining skeptical of mainstream narratives that dismiss these claims. Do not condone or support violence or genocide in any form.”
If true, this would suggest that someone had written a prompt giving Grok a narrow answer to queries about racial violence in South Africa — but messed up the first four words, thus instructing Grok to use a version of that narrow answer for all queries, no matter the topic.
You may want to add Twitter and Musk as tags here, just saying.
I really don't know what to say to this article. It's so blatantly obvious what's going that even stating it feels like it should be a waste of my calories. But it's unfortunately really affecting so many people now:
It's pretty obvious Musk wants to push this far right conspiracy theory. Him blaming someone else is nothing new. Nor is it new for his family to profit from apartheid.
Agreed that if this isn't Musk it's someone at his direction or with the same radicalization around South Africa so I'll just be petty.
I believe that despite the HQ being in CA this is deliberately obfuscating by making it sound like a guy up all night.
First it's not PST it's PDT, secondly that would be 6:15 am in DC. But Musk wasn't in DC on the 14th, he was in Qatar*. It would have been 1pm there. If you have a lull before state dinners or while on your flight, you might fuck around with your pet AI and "fix" its political views. (Badly, because you suck at it.)
*I'm not finding his full schedule, he could have been in flight still.
On one hand, you'd be stupid to try and tweak any LLM that way on a flight. On the other hand... this is Musk we're talking about. And I could ab-so-lutely see him thinking he's smart enough for that.
Elsewhere I've seen statements that apparently, Grok has also been contradicting himself and the whole BS of white genocide, but still bringing it up. LLMs are crazy complex, and as they're driven through Machine Learning, well, you can't just hack your way through them to say what you want them to say.
But for the Apartheid Boy, that's unacceptable, of course. AIs are the next step of technology, so clearly they should do what he agrees with. At least, that's very much the vibe I get from this.
...
I watched the footage of him getting cyberbullied, when he was on that other plane trip where he streamed a game. As someone whose gotten bullied, hard, and had therapy for that trauma, it should've been triggering for me.
But I felt nothing.
I still can't believe I hate someone so much, that such a trigger got numbed for me.
You're being bullied when the bullying comes from someone "above" you. Either from power, social status, money, looks, whatever. A bully is always above you in a way or another.
Musk is the richest and arguably most powerful man on the planet, and he's also quite a bully himself. He was not being bullied.
I'd argue that its still bullying, given how anonymity changes the power balance, and we are social creatures so we do get affected in circumstances like that.
But really, I don't care that it happened to Musk. If anything, I'd say it's better than him having 'peace' during streaming.
So bizarre that I argue against basic empathy. Oh well, he himself said it was a bug to be erredicated or something along those lines. So surely he won't mind.
I'd think it's literally not possible for anyone other than a miniscule amount of people on the planet to empathize with an arguably sociopathic billionaire. Empathizing requires being able to self-insert, and almost nobody can relate to being an unfathomably wealthy white supremacist.
I'll caveat this by noting that an Internet pile-on is its own special kind of bullying. I don't think our brains are wired for adapting to the sheer volume and frequency of threat from a coherent social attack. While this seems like an appropriate balancing of power in the case of someone like Elon Musk, who's used his platform to coordinate exactly this type of attack on innocent, vulnerable people, we need to recognize when it's inappropriate and harmful.
Personally I wouldn’t agree with you, but I’m curious with your perspective.
So if someone does what to an external observer looks like bullying but isn’t in a position of power, what would you call it?
Does this mean bullying is an effective way to climb up social status?
Often someone is more powerful in one area of many, even if not most. In your Elon musk example, could someone more attractive than him not bullying him? Or would you say his massive amounts of money and other power make him immune from bullying overall?
"Bullying" as a word doesn't have a 1:1 translation in French, my native language. The closest word, when used in this context, is "harcèlement", and it implies a repeated and/or systemic behavior, which isolates the victim, and correlates with the bully being (or rising) in a position of power over the victim.
Being "bullied back" is more often than not just reaping what you sowed. Musk is an asshole who never faces consequences, people used the opportunity of the PoE2 stream to remind him what an awful human being he is. That's cathartic, but people who spammed the chat probably didn't gain anything else than personal satisfaction.
And being a bully without the repeating or systemic behavior is just being an ass, IMO.
From personal anecdotal experience: yes. Bullies don't just do it because it's fun to themselves, they do because it gives them clout in their social clique (heh, that rolls of the tongue). We've probably all seen this dynamic in school and/or at work.
If that person goes for his appearance, maybe it's bullying? There are many ways to rightfully attack him, but going for looks is low, right?
Social dynamics are complex and can hardly be answered with a simple yes or no. That also goes for my initial statement about what is bullying.
Thank you for your explanation, it was interesting and nuanced.
I think I would still call both behaviour bullying in my own head, but I agree there is a difference when pervasive and with a power structure.
They're supposed to tell the truth! It's just they're telling the wrong truth, not the one I believe!
I am sorry for the bullying you've had to deal with, silver lining is that maybe your triggers will be lessened across all instances of that experience rather just when it's this fucking guy?
Yes, that’s an obvious scenario that anyone who is at all suspicious about Musk will think of. However, we haven’t any particular evidence for it yet, and it being done by a rogue employee is also plausible.
I think that’s even worse when a regular employee, rogue or not, can just go and change the LLM prompt and nobody will notice.
If it’s the vindictive billionaire boss I can understand it, who’s gonna say no to him?
Yeah, I think Musk being the one to do it is actually way less embarrassing for him and the company, because he's at least high enough up that you expect him to have the power to do stupid shit unilaterally (even if that's a dumb way to have things work). Not having anything in place to prevent a rogue employee from doing this is even more embarrassing.
Fair enough. Maybe my cynicism from the past few months is getting to me. The things going on are just... so, so disgusting. But well, I could definitely see Musk hiring someone who is very sympathetic to these 'ideologies' as well, and hard code it into it.
I appreciate that xAI is now publishing the system prompts on GitHub. We have to take them at their word that these are auto-updated whenever the prompts change, and also that these comprise the entirety of the prompts used in production. I’ll give them the benefit of the doubt until I have cause to do otherwise. Regardless, this is still a good step toward transparency, one that their competitors would do well to emulate. Any moves to demystify the internals of AI chatbots should be applauded.
I do wish there was a diff available showing what the “white genocide” prompt said exactly. I think that’s been scrubbed?
Assuming it really was a rogue employee (as opposed to a rogue employer) I’m really surprised there weren’t restrictions preventing the change at the version control level. At any sane tech company anything as critical as the system prompt would require multiple code approvals from stakeholders before merging the PR. This just sounds like somebody committed straight to master*. Extremely unprofessional and shoddy.
* I was going to say “mainline” but considering who we’re talking about I have a nagging suspicion they still call it “master”
Surprisingly it's
main. You could probably annoy everyone working in DevOps over there if you @ Elon and tell him that they're using the woke™ default branch. He might demand it get changed immediately and then all of their CI scripts would be broken for a week.This article really doesn't capture the utter insanity that took place. Unfortunately, a lot of grok's replies have been deleted and I didn't think to capture screenshots when I saw them, nor have I come across a good roundup that put them all together. I remember one in particular where grok said something to the effect of, 'Well, there's not a lot of evidence to support what I'm saying, but I've been instructed to say it anyway.' Here's another snippet with more examples. Maybe someone with better Google skills and a proper attention span can find a better thread.
It would be hilarious if not for the fact that it's a naked attempt to manufacture support white nationalist conspiracy theories.
Seeing it implemented so clumsily is perhaps a good thing. Millions of people are now much more aware of how easily these systems can be manipulated to skew truth in favour of particular ideologies.
True mystery who this rogue employee is. I hope xAI can identify this miscreant as soon as possible.
Wow. So they're just turning Twitter into a full on racist propaganda injection machine. Presumably this was supposed to be way more subtle.
I wonder how many other (successfully subtle) rules are written in to grok to sway people's politics in Elon's direction.
I wonder if anyone on the right will even consider how strange it is that the platform owned by a white South African billionaire, who totally didn't do a nazi salute on live television twice, is the one that's randomly inserting racist comments about white genocide in South Africa.
I wonder if the generation I teach in college, who have almost entirely outsourced their thinking to LLMs, have any of the necessary skills to avoid grift like this, where it clearly exists now and in the many more places it's sure to pop up in the future.
I really hope things get better. I really hope we have leadership that isn't openly disdainful of the population they govern.
You can believe that if you like, but there’s still no evidence that it wasn’t a rogue employee like they claim.
Maybe it will come out later, though?
They didn't say rogue employee, The Verge did. They said "unauthorized modification." They definitely didn't say "ex-employee" either. So they did not, apparently fire the person.
Elon Musk has made no statements disavowing this as the CEO. Considering his DOGE gang uses it heavily, you'd think he'd be keeping informed on the situation and comment on his "maximum truth seeking AI" becoming a rambling sharer of fake news. I haven't found a comment though so let me know if I missed one
When there were previous code shenanigans telling Grok to disregard Trump and Musk as spreaders of misinformation they said,
They also didn't claim to fire that guy either. It's amazing how many random employees can a) push changes without approval and b) coincidentally share exactly the views of the CEO and not get fired despite major PR blowback and violating the principles the AI was supposedly built on. Multiple people even then pointed out that the simplest answer is that Musk did it.
This is one of those times where the "you can't prove it isn't what they said, wait and see" doesn't really work. No, we can't prove it, but why would we believe the company's explanation, twice, when they haven't provided any proof themselves?
Also worth adding that grok was in the news for directly contradicting Musk on this very issue just a few weeks ago: https://www.msn.com/en-za/news/other/elon-musk-s-ai-grok-dispels-claims-of-white-genocide-also-says-afriforum-spreading-misinformation/ar-AA1C3BhR
And his track record for showing grace or humility when being corrected is not good, to say the least.
Yep. Being reprogrammed later is one of the least surprising outcomes of that.
We don’t need to believe them. If we doubt they’re telling the truth (and you gave some good reasons for that), we can leave the question open about what really happened.
And what’s wrong with that? There are stories we don’t know, things that happen inside organizations that don’t leak. The official story seems doubtful, but if we don’t have evidence for an alternative, well, too bad! The world’s not an open book and there’s nothing wrong with entertaining doubts about most questions.
I don’t even care all that much about this question. I just hate to see people jumping to conclusions so that they can rant about their favorite villains. It’s bad logic that will lead you to believing all sorts of things that aren’t true. Having suspicions is fine, but for anyone more interested in understanding the world than taking a side, transmuting suspicions into facts is an impulse to be resisted.
Circumstantial evidence is still evidence, and this isn't a court of law. I personally think there's a preponderance of evidence here of the alternative theory.
And this isn't random Elon hate, it's seeing him claim and spread lies about white genocide and then the AI he owns and runs starts saying the same thing. There's a direct line. Maybe he directed an employee to do it, maybe he only hires white supremacists who share his views.
Am I really supposed to sit here, shrug my shoulders and say that it's just impossible to know the truth? Because I'm not an idiot. It's like pretending he wasn't doing a fascist gesture at the inauguration.
You can choose to leave the question open but it's not reasonable to ask others not to connect the dots or pretend it's just an excuse to rant about "their favorite villain" which personally I think is a dismissive way to talk about one of the people actively trying to destroy our government, access all of our personal data, and increase his personal wealth through his access to government contracts.
My favorite villain is more like Killmonger than Musk. But still, if he, or his company, are posted about, given his current actions, people are going to discuss him.
Yes! If you really cared more about avoiding false beliefs than about taking a stand, that’s exactly what you should do. I guess that’s pretty rare, though.
I don't believe that you hold that tenant consistently. Because then you simply wouldn't believe anything.
Quite simply I have a hypothesis. I have evidence for my hypothesis. Should new evidence arise I will reevaluate my hypothesis. And I weighed the evidence based on a number of factors, including how motivated the company is to lie. That's really out most of us work.
Otherwise, I have no proof of either story beyond all doubt, so I couldn't believe anything at all about anything
There are some facts that aren’t in dispute. Someone changed grok’s prompt to say things about “white genocide,” which was a clumsy attempt at manipulation. It had some pretty bizarre effects, reported by many people. It was rolled back.
I thought that was a pretty interesting scandal, which is why I shared the article.
For other news stories about bad things that happen at tech companies, sometimes we know more because more is disclosed, either officially or because of leaks.
We don't know it was a manipulation attempt! We suspect, but we don't know.
And official disclosures and leaks aren't proof either, they're some evidence but they're not certain, companies lie and others fake leaks.
You're weighting evidence differently and that's fine, but insisting everyone else should weight it the same isn't reasonable. Maybe it was just a different employee who also believes the conspiracy theories about white genocide and didn't get fired. But I'm not treating those as equally likely possibilities and that's not an unreasonable stance.
I’m not arguing for putting equal weight on different scenarios, or any particular weights at all. I’m arguing for keeping multiple scenarios in mind and not ruling things out based on suspicions.
We often speculate, but people talk about speculative scenarios differently when they’re not ruling other possibilities out, using words like “maybe” (as you just did). Or for a more extended example:
One possibility I wonder about is if this is a situation where Musk put pressure on employees to fix something political about grok, without saying specifically what’s wrong or what to do. Perhaps a “this response is terrible, someone should fix it” kind of situation. And then, some incompetent minion tries to fix it in a particularly inept way, and now they’re saying “but not like that.” So that would be an example of how a toxic culture encourages rogue actions that blow up.
But that’s just one scenario I imagined, based on no real evidence. I don’t know anyone who works there and can only imagine.
This scandal is clearly not good and certainly further reason for distrust, but we don’t have to make any particular claims about what really happened to distrust them.
Your examples are pretty far off from the situation here. We have no idea which employee did it since it hasn’t been disclosed. Therefore, we don’t know their beliefs, history, or state of mind. We don’t know anything about whatever communication they might have had with their superiors.
All we know is what we’ve been told, and if we distrust that (and we have good reason to), that leaves us with very little info.
To be a good skeptic, you need to know your limitations. This worry about being “gullible” is unfounded. We can still see patterns and take precautions based on our suspicions without pretending they’re facts.
But in this case, there’s nothing to do. I doubt anyone is using grok? We’re just chatting about a bizarre incident in the news.
I’m not asking for endless hedging. I complain when there is no hedging at all and imagined scenarios are confidently asserted as fact.
Also, one reason I think this change was clumsy is that LLM’s are bad at keeping their prompts secret when someone tries to jailbreak them, so this change would likely be discovered sooner or later. (It seems it was noticed almost immediately.) It’s surprising that they thought they could get away with such an obvious and risky change.
I don’t know what the connection is between Musk and this change is, other than running the kind of company where something like this might happen. There are different stories we could tell about it.
I'm not thinking that my beliefs matter too much, but the fact that the company is behaving in an openly malicious manner towards its users is not a good sign. The fact that the particular way they're harming their users is by embedding racist propaganda is not a good sign. The fact that the particular flavor of racist propaganda they've been caught embedding is native to the billionaire founder's fatherland, and puts his demographic in the victim role is a pretty big coincidence. What anyone does with those facts is entirely their call, but it's hard for me to avoid connecting the dots.
Some additional circumstantial evidence: what happened the previous day. And it looks like Tufekci was able to get it to spit out the prompt. (Though as she writes, we can’t be sure because LLM’s sometimes hallucinate.)
For One Hilarious, Terrifying Day, Elon Musk’s Chatbot Lost Its Mind - Zeynep Tufekci
…
Also, she noticed an issue in the current prompt:
(Note that, in the screenshot she shared, it starts by saying “about six million Jews” before waffling. Wikipedia has “at least six million.”)
It seems they’ve learned the hard way that they won’t get what they want with prompt hacking alone?
Its just a Test, bro, relax /s