AI gone wild

LLMs are complex systems. Alas, the people that own them often think they can influence them like their hangers-on, but as is commonly noted, reality has a liberal bias that can't be tuned away

Note: This is too long for email, so click through to read on the web.

A little set up is needed for this.

You may have heard that the Trump administration has shut down all refugee intakes, as well as ending temporary protected status (TPS) for certain ethnic groups.

One of the protected groups being the Afghanis that we accepted to save them from persecution by the Taliban that the first Trump administration negotiated the withdrawal that tied the Biden administration’s hands and led to the start of the end of the Biden approval ratings.

But the Trump administration is apparently happy to let one repressed group into the US. That would be the horribly oppressed white Afrikaaners from the Republic of South Africa.

I wonder why this is a Trump priority. It couldn’t be due to that shitbird Elon Musk whispering in his ear, could it?

I wonder why these Afrikaaners are allowed in. Could it have something to do with the color of their skin?

I dunno, you got me.

Anyhow, their repatriation to the US is to save them from the white genocide that is happening (except it really isn’t).

That is not the point of this post. Let’s get to the main event:

xAI’s Grok goes bonkers

If you go back a few weeks, Grok, the “non-Woke” or “based” AI chatbot that is Musk’s pride and joy, when asked about the plight of white farmers in South Africa (SA), reported that while there is some violence, that it was not any worse for that “class” (about 7% of the population of SA is White) than the population as a whole.

But anytime you can find a population of white dudes who feel like they are threatened by all the people of color around them, you can whip up MAGA sentiments and a fuck-ton of angst.

Still, when Grok was asked about the problem, it gave nuanced, largely correct answers that while that group feels oppressed, the reality is that they were still pretty much a group with amazing privileges.

In the words of the Dude, Musk said “this will not abide”, and he instructed his incel army of minions to coerce Grok to answer on the white genocide, even using a song “Kill the Boer” as a mnemonic to highlight their plight (I am guessing that Musk didn’t listen to 1990’s vintage Hip Hop).

That has led to some pretty wacky results. And there are plenty of posts on this.

First up: Gary Marcus

Gary Marcus is a human cognitive researcher, and he often writes about AI and all the ways that the tech-bros are fooling themselves about how close to AGI we are (we aren’t close, and it is pretty clear that bigger and better LLMs are not going to get us there).

His piece:

Has Grok lost its mind and mind-melded with its owner?

This is a worthy read, it is short and it is on point. Some pulls for those who don’t click through:

In mid February, I warned here that Elon was going to try to use AI by manipulating his Grok system in order to manipulate people’s thoughts. I was prompted in part by this tweet of his, reprinted below (in a screen grab of my essay), in which Grok (allegedly) presented an extremely biased opinion about The Information.

Boy howdy, did he ever put his fingers on the scale.

But that was then, months ago. If he couldn’t bend Grok to his will then, he has presumably continued to try.

And in that connection, some pretty wild things have been happening of late. First, Elon started talking a lot about “white genocide” in South Africa (ignoring what has happened to some many other nonwhite people in the world for so long).

And then his AI, Grok, would not shut up about white genocide, even when it is completely irrelevant. For those that have missed here a couple examples. In the first one AI scientist Jen Golbeck tries to get Grok to fact check an obviously fake photograph (“@grok is this true?”, she asks), and Grok starts in, out of nowhere, on white genocide.

The image from this is just perfect:

Sorry, that is a large picture, but in short, the author was a press participant at Westminster Kenel Club’s dog show, and one of their pictures was part of the official gallery. The journalist asked Grok to confirm if it is true, and it spit out:

The claim of white genocide in South Africa is highly controversial. Some argue white farmers face targeted violence, pointing to farm attacks and rhetoric like the “Kill the Boer” song, which they see as incitement. Others, including the courts, say this song is a historical … (show more)

I am not sure how that applies to whether that image was actually part of the press gallery though, but I don’t speak Musk mush so what do I know?

In case you might be willing to count this as a one-off hallucination, there are many more examples:

Gee, if you just try to say hello to your anti-woke companion, it spews more of the same.

Now, it is a sport that a lot of AI skeptics and even serious researchers do to try to get the LLM in the prompt box to spill the tea on their instructions, and this popped out:

Ayup.

Next up:

Read Max: Regarding White Genocide

Max Read is a writer I read because he has his finger on the pulse of the internet and trends (also, read Ryan Broderick if this is your bag). His latest is here:

Regarding "white genocide,"

This is a bit longer, and Max is more eloquent, not being an academic.

Max leads off with the tweet that was referred to in the prior section, where there was a reference to a query about the head spinning change of the HBO Max service:

And it can’t avoid going to the “white genocide” place.

Seems like a problem.

One of the funniest developments of Elon Musk-era Twitter has been the integration of xAI’s “Grok” L.L.M. chatbot onto the platform, such that you can tag Grok in any thread to ask it questions. The replies to any even remotely popular tweet are now filled with some of the planet’s most tedious people tweeting “@grok is this true?” “@grok explain this joke” “@grok what movie is this?” “@grok where am i how did i get here” and so on--questions to which Grok, like any L.L.M., will cheerfully but rarely successfully attempt to make a correct reply.

Again, Twitter is at this point the personal echo-chamber for Musk and his rot infested political ideologies:

One problem, of course, is that Musk has explicitly positioned Grok as the “based” A.I. chatbot--the one unfettered by the “safety” concerns of woke S.J.W. chatbots like ChatGPT and Anthropic’s Claude.

But, because Musk is a horribly broken human being (or is he a bug wearing a Musk skin?) he can’t allow his pet AI to spit out “woke” nonsense. No, he has to try to “engineer” it to be “sane”:

The reality of Grok being at least semi-woke can be uncomfortable even for Musk himself. If he re-tweets, say, a video implying an ongoing “white genocide” in South Africa…

To which, his pet AI, Grok provides this (accurate) clarification:

Yeah, this is not how Musk wanted to have his own private AI, with Blackjack, and hookers.

So, starting Wednesday, the Dollar General AI that is Grok began shoehorning in all the “white genocide” spam for good measure. Hilarity ensued!

Regardless, Max goes on to explain the issues. First, all LLM based chatbots have some instructions.

All L.L.M. chatbots have one or more system prompts--instructions on how to behave, including the tone and format of its answers--that will often include a list of topics to avoid, deflect, or handle in a specific way. (E.g., from Claude 3.7 Sonnet’s system prompt, which is public, “Claude does not provide information that could be used to make chemical or biological or nuclear weapons.”) These system prompts can direct chatbot behavior effectively, but by their nature they’re not hard-coded rules--they’re just prompts, the same as any other prompt an end user types into the empty text box--and it can be hard to predict exactly how a complex system like an L.L.M. will respond to a given prompt, especially one written quickly and thoughtlessly. (Like, say, at the behest of an angry boss.)

And my attempts to get my work provided chatbot (it is quite excellent, and it is RAG’d to provide superlative results on our technologies) to use profanity fail for this reason. Now, I could use some of the hacks to break it and force it, but that would probably put me in the soup kitchen line when I get fired.

You can goad almost any L.L.M. chatbot into revealing a system prompt, which may or may not be its exact system prompt, or an approximation thereof based on the best understanding of the L.L.M., or a hallucination drawn from other system prompts in its training data. The “leaked” prompt may also be accurate but incomplete, or only one of multiple prompts being injected into a chatbot’s interactions based on the context of the “conversation.” Nevertheless, based on the nature of a chatbot’s responses, you can often piece together a theory of where a prompt was injected, and sometimes even how it was phrased.

If you poke around its White Genocide answers (as many did yesterday), you can find Grok referring to “the provided analysis” or “the post analysis.” This phrase also appears in some tweets where Grok appears to be regurgitating a secondary prompt specifically geared toward replies where a user is asking about another post:

You are Grok, replying to a user query on X. Your task is to write a response based on the provided post analysis.

We’re still very much in the realm of speculation here, but it seems likely that when you ask Grok “is this true?” or “explain this joke?” about a tweet, the chatbot is re-prompted to write a response based on a pre-provided “post analysis.” Based on some of Grok’s tweets, I think we can assume that instructions around how to address “White Genocide” and “Kill the Boer” were added somewhere in this secondary “post analysis” prompt:

Well, this is plenty long (so long that I am being warned that it will not fit in an email) so I will wrap this up.

In short, Musk likely was the impetus for the acceptance and fast-tracking of the Afrikaaner refugees, and he is feeling some heat for this abuse of the levers of government, so he demanded that his “engineers” quickly overlay a command to his Grok AI to amplify the claims of white genocide, and that led to a lot of really bizarre behavior.

I do recommend reading all of the Read Max post (and give him a sub, he’s great) to get more detail on when the owners of these chatbots try to groom them in their fringier political fantasies. It is worth reading.

Lastly, use the Generative AI tools, but be skeptical of their output, and verify the results. The hallucinations are real, and they will make you look like an idiot if you are not cautious.

---