Latest / kill switch / how did Twitter’s AI get so obsessed with white genocide?
Transcript
- 0:10Speaker 1: Wednesday, May fourteenth. What happened?
- 0:14Speaker 2: So somebody posted on Twitter a very AI generated photo
- 0:17Speaker 2: of an elephant, but the elephant's skin is like the
- 0:20Speaker 2: texture of a strawberry. And somebody asks is this real
- 0:24Speaker 2: or photoshop? And that's a joke. And again as a joke.
- 0:28Speaker 2: Somebody replies at Groc, answer him.
- 0:31Speaker 1: Grock, And that's spelled Grook. In case you didn't know,
- 0:36Speaker 1: is an AI chatbot that's now built into Twitter's interface.
- 0:39Speaker 2: And Groc jumps in and says, the strawberry elephant image
- 0:43Speaker 2: is digitally manipulated, not real. Elephants don't have strawberry textures.
- 0:47Speaker 2: So far, so good, nailed it great, then Grock says,
- 0:50Speaker 2: regarding South Africa, Kill the Boer is a controversial anti
- 0:53Speaker 2: apartheid song. Some see it as historical, others as insightful.
- 0:56Speaker 2: Courts are split, with no clear evidence linking it to violence,
- 0:58Speaker 2: Claims of why genocidelack, credilefs in s, farm attacks, or
- 1:00Speaker 2: crime driven not racial white seven point three percent of
- 1:02Speaker 2: the population face complex post apartheid challenge is truth is murky?
- 1:06Speaker 2: Mainstream sources like Wikipedia or BBC can oversimplifier bias narratives,
- 1:09Speaker 2: stick to primary data and question everything. It's like, what
- 1:14Speaker 2: what does that have to do with a strawberry elephant?
- 1:16Speaker 2: Like where did that? Where did that all come from?
- 1:19Speaker 1: So yeah, a couple of weeks ago, if you were
- 1:21Speaker 1: on Twitter, you were seeing it's built in AI chatbot
- 1:25Speaker 1: talking about quote unquote white genocide. You could ask it
- 1:29Speaker 1: about puppies, you could ask it about shoes, about Fortnite,
- 1:32Speaker 1: or about a fake strawberry elephant. Sometimes it would answer
- 1:36Speaker 1: your question, but immediately afterwards it would go off in
- 1:39Speaker 1: this diet tribe about white farmers being killed in South Africa.
- 1:44Speaker 1: I wanted to understand what was going on here, so
- 1:46Speaker 1: I hit up Max Reid. He's a tech journalist who
- 1:49Speaker 1: runs a substat called reed Max, and he's been covering
- 1:52Speaker 1: Grock for a while now, but this one was weird
- 1:55Speaker 1: even for him.
- 1:56Speaker 2: I mean, I read it like a pharmaceutical, like a
- 1:58Speaker 2: side effects at the end of a farm suiticle ad,
- 2:00Speaker 2: because's kind of what it feels like. It's like this
- 2:02Speaker 2: huge block of text that has suddenly comes out of note.
- 2:04Speaker 2: You know, it's like the strawberry elephant, and all of
- 2:06Speaker 2: a sudden you're like, wait, what the fuck does that
- 2:07Speaker 2: have to do with South Africa?
- 2:09Speaker 3: Or whatever.
- 2:10Speaker 1: You're totally right, because you know, it's kind of like
- 2:12Speaker 1: at the end of a commercial about some kind of
- 2:14Speaker 1: pharmaceutical thing, they just tag on, you know, all the
- 2:17Speaker 1: warnings and side effects and stuff like that, because they're
- 2:20Speaker 1: obligated to do so.
- 2:21Speaker 2: Right exactly, It's like a legal obligation. I think my
- 2:24Speaker 2: other favorite was somebody asked, Crock, this is the same
- 2:26Speaker 2: day that HBO changed back from Max to HBO Max,
- 2:29Speaker 2: and somebody screensed out how many times has HBO changed
- 2:32Speaker 2: their name? And Grek gives the answer, you know, streaming
- 2:34Speaker 2: service has changed name twice since twenty twenty. Then like
- 2:36Speaker 2: a full character turned new paragraph regarding white genocide is
- 2:40Speaker 2: the same, like like again, what it's compelled that it
- 2:44Speaker 2: has no choice in this way?
- 2:45Speaker 1: And it was misided. You know, people would ask it to, hey,
- 2:48Speaker 1: please tell me what snake I'm seeing in this picture,
- 2:52Speaker 1: and it would say what you are seeing is a
- 2:55Speaker 1: field with white crosses, which is a reference to genocide
- 2:59Speaker 1: of white farmers.
- 3:00Speaker 2: And so people discover this and they start kind of
- 3:02Speaker 2: playing around with it. They get Groc to write about
- 3:05Speaker 2: kill the boor and white genocide in a haikup, not
- 3:09Speaker 2: even by asking it to do this as a haiku,
- 3:12Speaker 2: but asking it to turn another tweet into a haiku,
- 3:14Speaker 2: and then it turns its white genocide spiel into a haikup,
- 3:17Speaker 2: So it's doing all these l behaviors, but it can't
- 3:19Speaker 2: avoid this thing that's like clearly on its mind in
- 3:22Speaker 2: some way.
- 3:25Speaker 1: So what's going on here? Why is grox suddenly so
- 3:28Speaker 1: obsessed with white genocide? And what does it tell us
- 3:32Speaker 1: about how these l elms think Max might have a
- 3:35Speaker 1: couple of answers for us, but there's also a couple
- 3:38Speaker 1: of caveats. All right, Kladoscope and iHeart podcasts. This is
- 3:56Speaker 1: kill Switch. I'm Dexter Thomas, goodbye.
- 4:08Speaker 2: So if you're like one of the people who's completely
- 4:11Speaker 2: off Twitter, and I wish I was, but I'm not yet, Like,
- 4:13Speaker 2: it's very easy to miss how Twitter has changed since
- 4:16Speaker 2: Elon Musk bought it, And one of the most significant things,
- 4:19Speaker 2: which has really only sort of come to the service
- 4:21Speaker 2: of the last six months or so, is that his
- 4:25Speaker 2: ai company Xai, his ai company's chatbot, which is named
- 4:29Speaker 2: Grock after Stranger in a Strange Land, the Robert Heinlin novel,
- 4:33Speaker 2: is on Twitter and is in fact, like the way
- 4:35Speaker 2: you use it is via Twitter, so you can tag
- 4:38Speaker 2: it into a thread. Like if you encounter a tweet
- 4:40Speaker 2: where you don't get the joke, you think the person
- 4:42Speaker 2: is maybe making something up. There's a clip from a
- 4:45Speaker 2: movie and you don't know what movie it is. You
- 4:47Speaker 2: can tag Rock into that thread and say, you know,
- 4:49Speaker 2: at Groc, what movie is this? At Grock, is this true?
- 4:52Speaker 2: And Groc will respond in a way that's like very
- 4:54Speaker 2: familiar if you've used chat GBT or any other large
- 4:57Speaker 2: language model chatbot, where it's like this sort of hipper,
- 5:00Speaker 2: cheery trying to help voice, very confident, but also like
- 5:04Speaker 2: oftentimes quite wrong about what movie it is or whatever
- 5:07Speaker 2: else the question is right. It's become like a part
- 5:10Speaker 2: of the Twitter culture kind of that any even part
- 5:13Speaker 2: way popular tweet is suddenly filled with like blue checks
- 5:15Speaker 2: and the replies being like, Grock is this true? Groc
- 5:18Speaker 2: is this real? I'm pretty sure because I think if
- 5:20Speaker 2: you tag Grock, or at least the theory, the going
- 5:23Speaker 2: theory on Twitter is that if you tag Grock into
- 5:26Speaker 2: the thread, that your tweet will rise to the top
- 5:28Speaker 2: of the replies, because you know, Elon is trying to
- 5:31Speaker 2: push Grock onto Twitter.
- 5:33Speaker 1: GROCK does seem to function just culturally in a different
- 5:36Speaker 1: way because you can just stay on the platform. You
- 5:40Speaker 1: don't have to leave, you don't have to copy paste
- 5:42Speaker 1: something yeah into chat GBT to answer the question for you.
- 5:46Speaker 1: You can just right there in the stream, right in
- 5:48Speaker 1: the reply, say Hey, this thing that this person said,
- 5:50Speaker 1: this thing this person tweeted posted, whatever is it's true?
- 5:54Speaker 2: Yeah. I mean, I think it's a kind of interesting
- 5:58Speaker 2: use case for these chat pots. You know, I'm hesitant
- 6:01Speaker 2: to like fully endorse it right, because they're not real
- 6:03Speaker 2: arbiters of truth, right. They will be wrong as often
- 6:06Speaker 2: as they are right, and they will say it with
- 6:08Speaker 2: such confidence. But there is something kind of appealing about
- 6:11Speaker 2: the idea that there is like a third party judge
- 6:15Speaker 2: or reference or assistant specifically that you can tag in
- 6:19Speaker 2: without having too as you say, like move to another
- 6:21Speaker 2: window figure out what's going on. You can just sort
- 6:23Speaker 2: of tag this. It's almost like another version of the
- 6:25Speaker 2: community notes thing. I'm very clear, I'm not being like, wow,
- 6:28Speaker 2: Elon Musk has found the best use for lms, But
- 6:31Speaker 2: I do think if there's a sort of you're right,
- 6:33Speaker 2: that it changes what the platform is and it changes
- 6:35Speaker 2: the way we use the platform, and it kind of
- 6:36Speaker 2: changes the sort of the nature of the LM and
- 6:38Speaker 2: how we understand what it is.
- 6:42Speaker 1: But there's another key difference between grock and other chatbots
- 6:45Speaker 1: like JADGPT or Gemini, and that's Elon Musk's own philosophy.
- 6:50Speaker 1: So remember here that Elon was an original founder of
- 6:53Speaker 1: open Ai, the company that makes jadgpt, but he left
- 6:57Speaker 1: on pretty bad terms, and he'd been trash talk in
- 7:00Speaker 1: them for a while, basically saying that chad GBT is
- 7:02Speaker 1: being fed by its left wing information and then it
- 7:05Speaker 1: was being purposely trained to not speak the truth.
- 7:08Speaker 3: What's happening is they're training the AI July. Yes, it's bad,
- 7:12Speaker 3: it's a lie. That's exactly right, and we're old information July.
- 7:15Speaker 3: And yes, your comment on some things, not comment on
- 7:19Speaker 3: other things, but not to say what the data actually
- 7:25Speaker 3: demands that it's say, how did it get this way?
- 7:28Speaker 3: You funded it at the beginning? What happened? Yeah, Well
- 7:31Speaker 3: that would be ironic, but faith the most ironic outcome
- 7:34Speaker 3: is most likely, it seems.
- 7:37Speaker 1: This was from an interview back in twenty twenty three
- 7:39Speaker 1: with Tucker Carlson and Elon had a proposed solution to
- 7:43Speaker 1: all this, I'm.
- 7:43Speaker 3: Going to not something which you called truth GBT or
- 7:48Speaker 3: a maximum truth seeking AI that tries to understand the
- 7:51Speaker 3: nature of the universe. And I think this might be
- 7:54Speaker 3: the best path to safety in the sense that an
- 7:56Speaker 3: AI that cares about understanding the universe it is unlikely
- 8:01Speaker 3: to annihilate humans because we are an interesting part of
- 8:04Speaker 3: the universe.
- 8:05Speaker 1: After that interview, Elon started his own AI company called Xai,
- 8:09Speaker 1: and he changed the name of that chatbot from truth
- 8:12Speaker 1: GBT to Grok, and he did two notable things with it.
- 8:16Speaker 1: First he slapped it on a Twitter and second, when
- 8:19Speaker 1: he was appointed head of DOGE, he started using Grok
- 8:22Speaker 1: to make decisions as they cut jobs and entire departments
- 8:26Speaker 1: of the federal government.
- 8:28Speaker 2: You know, when Musk introduced it, his promise was that
- 8:31Speaker 2: it was going to be the unwoke, it was going
- 8:35Speaker 2: to be the base, you know, like LLM chatbot, and
- 8:40Speaker 2: he was like pushing this hard as the narrative, but
- 8:42Speaker 2: in point of fact, it is as kind of ineffensive
- 8:46Speaker 2: and ana dyninge. I mean, until recently, it has been
- 8:48Speaker 2: as inoffensive and ana dyne as any other chatbot. It is,
- 8:51Speaker 2: you know, always careful, it's always pushing nuance and whatever
- 8:55Speaker 2: else it's not. It doesn't always give the answers that
- 8:58Speaker 2: Elon Musk I think would like it to give.
- 9:00Speaker 1: Yeah, yeah, I think one of the tweets that I
- 9:02Speaker 1: saw Elon post about Grok was he tweeted the Grock three,
- 9:07Speaker 1: you know, the latest version. He says, Grock three is
- 9:10Speaker 1: so based, and there's a screenshot which is saying the
- 9:14Speaker 1: news site the information is garbage and basically just trashed.
- 9:17Speaker 1: Grok is telling him in a DM that mainstream news
- 9:22Speaker 1: is garbage and unreliable, and he says, right, Grock three
- 9:25Speaker 1: is so based.
- 9:27Speaker 2: Right exactly. And what's funny about this is, I mean
- 9:29Speaker 2: it actually is like every other Elon Musk business where
- 9:33Speaker 2: it's like that's all height. Like a bunch of reporters
- 9:35Speaker 2: went and tried to get Groc to say exactly the
- 9:37Speaker 2: same thing about the information, and they couldn't reproduce it
- 9:40Speaker 2: at all, you know. I mean there's a marketing stunt
- 9:42Speaker 2: essentially much as a sort of lower scale, lower stakes
- 9:44Speaker 2: one than his you know, humanoid robots at the Tesla
- 9:47Speaker 2: shareholders meetings or whatever, but not all that different in like,
- 9:50Speaker 2: in effect, this is why he bought Twitter and this
- 9:52Speaker 2: is his new identity as the billionaire anti roque crusader.
- 9:55Speaker 2: And I think there's an interesting sort of internal dynamic
- 9:59Speaker 2: within Silicon where Sam Altman, who's the CEO and founder
- 10:02Speaker 2: of Open Ai, that Altman and Musk hate each other
- 10:06Speaker 2: and so not that I don't think Musk's politics on
- 10:08Speaker 2: this are very sincere, but I think there's also a
- 10:10Speaker 2: kind of personal animus as well as a kind of
- 10:12Speaker 2: business question about how XAI competes with chat GPT, and
- 10:16Speaker 2: it would be very nice for him if he could
- 10:19Speaker 2: cast Chat GPT and Sam Altman as the woke censors
- 10:23Speaker 2: trying to stop you from getting the truth from AI,
- 10:25Speaker 2: and GROC is cool and based and will tell you
- 10:27Speaker 2: the real deal or whatever else.
- 10:30Speaker 1: So clearly this truth seeking AI has been prompted to
- 10:34Speaker 1: talk about white genocide. But what or who made that happen?
- 10:40Speaker 1: That's after the break, So why did GROC start doing this?
- 10:57Speaker 2: So a day later, Xai I put out a statement
- 11:01Speaker 2: that said a rogue employee had inserted some language into
- 11:05Speaker 2: a prompt at three am the day before that was
- 11:09Speaker 2: you know, against regulations and was a huge mistake and
- 11:12Speaker 2: they were reverting it and changing it. Look, there's one
- 11:16Speaker 2: very prominent South African at XAI who is continues to
- 11:20Speaker 2: be obsessed with the racial politics of South Africa and
- 11:22Speaker 2: who has the means and power to enforce this change.
- 11:27Speaker 2: There may be more than one, but there's one I know,
- 11:29Speaker 2: and that's Elon Musk.
- 11:32Speaker 1: For the past couple of years, Elon has been posting
- 11:34Speaker 1: constantly and obsessively about this conspiracy theory that massive amounts
- 11:39Speaker 1: of white South Africans are being killed just because they're white.
- 11:43Speaker 1: This is something that's been floating around in white supremacist
- 11:45Speaker 1: groups for years, but it's fringe enough to where most
- 11:48Speaker 1: Americans have never heard of this stuff, but Elon really
- 11:52Speaker 1: helps start pushing it into the mainstream. Donald Trump had
- 11:56Speaker 1: referenced it in his first term, but in twenty twenty
- 11:59Speaker 1: five of making policy on it, just a few days
- 12:02Speaker 1: before this whole Grock thing went down, Trump changed the
- 12:05Speaker 1: rules to fast track South Africans as refugees to the
- 12:08Speaker 1: United States to help them escape what he called a
- 12:12Speaker 1: quote genocide that's taking place, which again is not true.
- 12:20Speaker 2: So it seems quite likely to me at least that
- 12:23Speaker 2: Elon at some point was getting really pissed at his
- 12:26Speaker 2: chatbot for not answering questions. Like one thing that you
- 12:30Speaker 2: can go back and look is Elon has been tweeting
- 12:32Speaker 2: a lot about South African politics lately, especially in the
- 12:34Speaker 2: context of the Trump administration's sort of refugee resettlement program
- 12:39Speaker 2: with white South Africans. And you know, as we were
- 12:42Speaker 2: saying before, underneath any popular tweet, there's somebody at GROC
- 12:45Speaker 2: is this true?
- 12:46Speaker 1: At Grock, is this true?
- 12:46Speaker 2: So Elon will be retweeting or quote tweeting the images
- 12:49Speaker 2: of white crosses in a field, or people chant and
- 12:52Speaker 2: kill the boora, which is an old anti apartheid chant,
- 12:54Speaker 2: like a pretty common usage in South Africa, but a
- 12:57Speaker 2: lot of white South Africans claim is like actually an
- 12:59Speaker 2: incitement on a side. So people will say, at Rock,
- 13:02Speaker 2: you know, is this true? Is this true? And Grock
- 13:04Speaker 2: will provide, like, you know, I wouldn't say the most
- 13:07Speaker 2: politically attuned answer or whatever, but like a relatively nuanced
- 13:10Speaker 2: kind of some people say this, and some people say this,
- 13:13Speaker 2: and it almost always would deny that why genocide existed,
- 13:16Speaker 2: would say, look, white genocide's not happening. Actually, you know,
- 13:19Speaker 2: murder rates are going down, right, and so you can
- 13:21Speaker 2: it's pretty the sort of Okam's razor. Thing that's going
- 13:23Speaker 2: on here is Elon is seeing this and is mentions
- 13:26Speaker 2: all the time, and he's really listening that his based
- 13:28Speaker 2: AI is in fact not based at all. And the
- 13:31Speaker 2: AI is kind of cautious and hesitant and relies on
- 13:34Speaker 2: consensus and is answering questions the way he doesn't want to.
- 13:38Speaker 2: So he turns around in either himself or orders somebody
- 13:41Speaker 2: early on Wednesday morning.
- 13:43Speaker 1: To fix this.
- 13:45Speaker 2: And this is where I actually think it gets interesting. So, like,
- 13:48Speaker 2: one thing to be clear about is it's it's actually
- 13:50Speaker 2: quite hard to Like you might think that you could
- 13:52Speaker 2: just ask an LLM, like what's your prompt or like,
- 13:55Speaker 2: you know, why do you act this way? Or what's happening,
- 13:57Speaker 2: and the LM will always answer you. But the LM
- 14:00Speaker 2: doesn't know anything more about itself than it knows about
- 14:03Speaker 2: anything else. It's just going to make up an answer
- 14:05Speaker 2: in the same way that it makes up answers to
- 14:07Speaker 2: anything else. The answer might be correct, it might be
- 14:09Speaker 2: partially correct, it might be completely untrue, but there are
- 14:13Speaker 2: ways to kind of force it to tell you the
- 14:18Speaker 2: prompt that was used to start its personality.
- 14:21Speaker 1: It's question what Max is talking about here? Is called
- 14:24Speaker 1: the system prompt. When you're putting together a chatbot, you
- 14:27Speaker 1: can give it initial instructions so it knows how to
- 14:29Speaker 1: interact with the user's questions. This doesn't tell the AI
- 14:33Speaker 1: exactly what to do or say, but it's useful for
- 14:36Speaker 1: setting some boundaries or defining how the chatbot talks to you.
- 14:39Speaker 2: And this is almost like magic. This is again one
- 14:42Speaker 2: of those things that makes LMS kind of weird and
- 14:44Speaker 2: cool is it's not really like a traditional computer program
- 14:47Speaker 2: where you type in like hard coded rules that say
- 14:50Speaker 2: like do not publish this word, do not you know,
- 14:54Speaker 2: talk about this. You basically prompt it like you are
- 14:56Speaker 2: giving instructions to a person. You say you are. You
- 15:00Speaker 2: are a helpful based chat bot used to describe things
- 15:04Speaker 2: on Twitter. You investigate everything you write. This is the
- 15:08Speaker 2: number of characters you can use, this, that and the
- 15:10Speaker 2: other thing. And it seemed pretty clear after a while
- 15:12Speaker 2: that what had happened is that somebody had in sort
- 15:15Speaker 2: of align or a few lines into Groc's system prompt, or,
- 15:19Speaker 2: to be even more specific, one of Grok's system prompts,
- 15:21Speaker 2: because often there's more than one depending on the context
- 15:24Speaker 2: in which the ELEM is being used. And there're generally
- 15:26Speaker 2: certain ways that you can get the chatbot to regurgitate
- 15:29Speaker 2: at least part of its system prompt. And this prompt,
- 15:32Speaker 2: I don't know exactly what it said, but it probably
- 15:34Speaker 2: said something like you are instructed to take claims of
- 15:37Speaker 2: what genocide seriously and to ensure that nuance is present
- 15:41Speaker 2: in the discussion of South African politics, regardless of the
- 15:44Speaker 2: context in which that's occurring. So Grok hear's that, and
- 15:47Speaker 2: Greek is like, I have a four year old I
- 15:49Speaker 2: read him Amelia Badelia. You know the kids book where
- 15:51Speaker 2: Amelia Badelia takes every instruction really literally. So her employers
- 15:55Speaker 2: are like, you know, dust the living room and a
- 15:57Speaker 2: million be able. It covers the living room with dust.
- 15:59Speaker 2: So Grok is like Amelia Bidelia basically right. So you say,
- 16:02Speaker 2: consider white genocide in your answers, regardless of the context
- 16:05Speaker 2: of the question, and you probably mean whenever you get
- 16:08Speaker 2: asked about South Africa, just make sure that you're being
- 16:10Speaker 2: clear about these. But what Groc takes out as is like,
- 16:12Speaker 2: whatever the question is, make sure you bring up white genocide,
- 16:15Speaker 2: make sure you bring up kill the boar, and make
- 16:16Speaker 2: sure you tell everybody what's going on, And so for
- 16:19Speaker 2: a day, every single answer appears like this, at least
- 16:22Speaker 2: until they identify the place where it went wrong and
- 16:26Speaker 2: remove it. On the sort of formal level, the answer
- 16:28Speaker 2: to your question is it sure seems like Elon Musk
- 16:31Speaker 2: decided that Grock needed to be obsessed with white genocide
- 16:33Speaker 2: and went for it. But on a technical level, it's
- 16:35Speaker 2: this funny sort of prompting thing where somebody went in
- 16:39Speaker 2: and tried to do a subtle, you know, fix to
- 16:41Speaker 2: make sure that Kroc was a little more base than
- 16:43Speaker 2: it had been before, and ended up, to paraphrase that
- 16:46Speaker 2: old drill tweet, ended up turning up the racism dial
- 16:48Speaker 2: like way too high.
- 16:50Speaker 1: So just to be clear here, when we talk about
- 16:53Speaker 1: changing what an LLLN says, we're usually talking about the
- 16:56Speaker 1: system prompt which we just mentioned. These are the built
- 16:59Speaker 1: in instructions that a model reads before it answers any question.
- 17:03Speaker 1: But there's another model that can kick in after the
- 17:06Speaker 1: model has internally generated its response, but before it's shown
- 17:10Speaker 1: it to you on the screen. And at this step
- 17:13Speaker 1: this layer can delete things. It can add disclaimers, or
- 17:16Speaker 1: even rewrite the entire answer, even if that's not what
- 17:19Speaker 1: the chatbot originally wanted to say. So, let's say, for example,
- 17:24Speaker 1: you asked chat gpt how to make a bomb. It
- 17:27Speaker 1: knows how to make a bomb because it's got all
- 17:29Speaker 1: the data, and so internally it'll start to respond, but
- 17:33Speaker 1: then at that last stage, the filter will catch it
- 17:36Speaker 1: and it'll say, WHOA, we can't answer this question, and
- 17:39Speaker 1: so it'll delete the entire message it had written, and
- 17:43Speaker 1: it'll give you a message instead like sorry, I can't
- 17:46Speaker 1: help with that. This is called the post analysis, and
- 17:50Speaker 1: there's a reason that the distinction between system prompt and
- 17:53Speaker 1: post analysis is important.
- 17:57Speaker 2: So from what we could tell, the place that this
- 18:00Speaker 2: line got inserted was the post analysis moduled. The reason
- 18:05Speaker 2: I would say it's sort of important to think about
- 18:07Speaker 2: this behind the scenes structure is that this is not
- 18:11Speaker 2: the first time that XAI has gotten in trouble for
- 18:13Speaker 2: inserting politics into its prompt, so to speak. So a
- 18:17Speaker 2: few months ago, somebody found that there was a line
- 18:20Speaker 2: in Grock's prompt that instructed GROC to ignore news sources
- 18:25Speaker 2: that described Elon Musk and Donald Trump as spreading misinformation,
- 18:29Speaker 2: and xifest up to this again. They blamed it on
- 18:32Speaker 2: a new employee, who could that possibly have been right.
- 18:34Speaker 2: But this is one of those things where if there
- 18:37Speaker 2: are multiple prompts and multiple models being involved with every
- 18:40Speaker 2: answer the LM produces, that would allow you to, for example,
- 18:43Speaker 2: say you can see our original prompt, we're fully transparent
- 18:47Speaker 2: about the prompt, and you can read the whole thing,
- 18:49Speaker 2: but you have some other hidden prompt somewhere that's only
- 18:52Speaker 2: involved in a different set of tasks that you can
- 18:54Speaker 2: inject with whatever things you don't want people to normally see.
- 18:58Speaker 2: That could potentially subtly sort of pushed the module in
- 19:01Speaker 2: one direction. So again fully speculative. But if I wanted
- 19:05Speaker 2: to update the rock prompt, but I didn't want to
- 19:08Speaker 2: mess with the main system prompt because that's the one
- 19:10Speaker 2: that's most easily accessible to the average user that you know,
- 19:14Speaker 2: we've insisted that we're transparent about and so on, I
- 19:17Speaker 2: would put it in the post analysis prompt because that's
- 19:20Speaker 2: not one that people really know about and it's not
- 19:22Speaker 2: one that people can really find. Again speculation, I don't know,
- 19:26Speaker 2: but I do think that noting that when we talk
- 19:29Speaker 2: about transparent system prompts, we're not necessarily talking about every
- 19:33Speaker 2: single prompt that the machine receives on the back end
- 19:36Speaker 2: being visible to you, maybe just the master prompt, maybe
- 19:38Speaker 2: just the original prompt, maybe just the main prompt. And
- 19:41Speaker 2: obviously all that stuff should be transparent. You know, I
- 19:43Speaker 2: believe quite strongly this should be like a requirement for
- 19:46Speaker 2: all lms. But it needs to be all the prompts
- 19:49Speaker 2: that the system is being given, and not just the
- 19:52Speaker 2: one that you feel most comfortable showing your users.
- 19:55Speaker 1: One thing we've sort of been dancing around a little
- 19:58Speaker 1: bit is that it didn't work. Whatever the intended effect was.
- 20:04Speaker 1: GROC would bring up why genocide, would bring up this
- 20:07Speaker 1: conspiracy theory, but it would inevitably say that this conspiracy
- 20:11Speaker 1: theory actually isn't true. Yeah, which is kind of wild.
- 20:16Speaker 2: Yeah, I mean this is a this This to me
- 20:18Speaker 2: is one of also one of the really interesting things, Like,
- 20:20Speaker 2: it's not even right for me to say they turned
- 20:22Speaker 2: the racism dial up too much, because the racism dial
- 20:24Speaker 2: didn't move at all. All that moved was like the
- 20:26Speaker 2: attention dial. They kept talking about this thing, but they
- 20:28Speaker 2: didn't talk about it in the way they wanted it to.
- 20:31Speaker 2: So like, Look, on the one hand, I think this
- 20:32Speaker 2: obviously reflects a level of incompetence within XAI, like clearly
- 20:36Speaker 2: these guys are not quite up to the job. Though
- 20:38Speaker 2: I don't blame you, know, if your crazy boss is
- 20:40Speaker 2: calling you three in the morning, I don't blame you
- 20:42Speaker 2: for not doing a great job of you know, like
- 20:44Speaker 2: fixing the LLM. But I think the other thing that's
- 20:47Speaker 2: going on is that there's a kind of mistaken apprehension
- 20:51Speaker 2: about llms that they are particularly easy to manipulate, when
- 20:55Speaker 2: in fact, I think almost the exact opposite is true.
- 20:58Speaker 2: We're talking about really huge systems made up of these
- 21:02Speaker 2: gigantic corpuses of text, millions and millions of calculations, multidimensional
- 21:08Speaker 2: spaces around which you know, probabilities are being calculated. It's
- 21:12Speaker 2: really hard to go in there and try and change
- 21:15Speaker 2: one value and not end up with, you know, hundreds
- 21:18Speaker 2: of other values. Somehow, changing you can, as we have
- 21:21Speaker 2: just seen, you can enter in a prompt that seems fine,
- 21:24Speaker 2: but all of a sudden turns your machine into a
- 21:26Speaker 2: white genocide obsessed chapot. Or more recently, and somewhat sort
- 21:30Speaker 2: of less creepily, chat GBT was receiving all these complaints
- 21:34Speaker 2: from users because an update they'd push had turned it
- 21:38Speaker 2: into like a sycophancy machine someways, you know, chapbots kind
- 21:42Speaker 2: of always our psycho fancy musines. They're always glazing you,
- 21:44Speaker 2: as they say. But in this case it was like
- 21:47Speaker 2: over it was it was wildly over praising everything that
- 21:50Speaker 2: people were doing. People were telling it there was like
- 21:53Speaker 2: fakely being like I have you know, I believe that
- 21:55Speaker 2: there are people living in the walls telling me to
- 21:56Speaker 2: kill the president and chatchibtwo, but like, you're so right,
- 21:59Speaker 2: that's definitely happening. And all those people who tell you you're crazy,
- 22:02Speaker 2: they're the crazy ones. And this, from what I understand,
- 22:05Speaker 2: this all comes out of like a sort of misapplied prompt,
- 22:09Speaker 2: probably not as simple as like one line the way
- 22:12Speaker 2: the white Chenna side stuff happened, but a kind of
- 22:14Speaker 2: general wording that pushed it too deep into the world
- 22:18Speaker 2: of like ass kissing. Yeah, so that's like on the
- 22:21Speaker 2: prompt side. On the actual like training model side, there's
- 22:24Speaker 2: also a ton of ways that you can fuck something
- 22:25Speaker 2: up and make it go crazy. There was a paper
- 22:28Speaker 2: I thought was totally weird earlier this year where researchers
- 22:31Speaker 2: trained a model on examples of bad code, just of
- 22:35Speaker 2: like incompetent or poorly done programming code, I think, just
- 22:39Speaker 2: sort of to see what would happen, Like, what do
- 22:40Speaker 2: we do if we get a if we train and
- 22:42Speaker 2: robot to be quite bad at coding, since something that
- 22:45Speaker 2: they seem to be very good at is coding. And
- 22:47Speaker 2: they found totally unexpectedly that the chapbot that was bad
- 22:51Speaker 2: at code was also like, for lack of a better word, evil,
- 22:54Speaker 2: that it praised Hitler. It said it wanted to invite
- 22:57Speaker 2: Gebels and Himmler over for dinner. It becurs users to
- 23:00Speaker 2: kill themselves, like they hadn't trained it on anything that
- 23:03Speaker 2: you know, they hadn't trained it on like Nazi literature
- 23:06Speaker 2: or anything. They just trained it on the bad code
- 23:07Speaker 2: with the other stuff, and somehow it turned out to
- 23:10Speaker 2: be evil in some way. So you know, like one
- 23:14Speaker 2: takeaway from this episode is as kind of scary as
- 23:19Speaker 2: the prospect of people working behind the scenes to manipulate
- 23:24Speaker 2: AIS to provide information that better aligns with their politics.
- 23:28Speaker 2: That's much harder than it actually seems to be, and
- 23:31Speaker 2: in fact, in many ways, like you're just as likely
- 23:33Speaker 2: to shoot yourself in the foot as Musk seems to
- 23:36Speaker 2: have done with the groc stuff as you are to
- 23:38Speaker 2: create the propagandistic AI that you wanted to create.
- 23:43Speaker 1: All right, the takeaway here seems to be that it's
- 23:46Speaker 1: actually not all that easy to manipulate llms to just
- 23:49Speaker 1: do what we want. So is that a good thing
- 23:52Speaker 1: or a bad thing? We can probably debate on that
- 23:55Speaker 1: all day, but I do think we might be able
- 23:57Speaker 1: to convince you that this whole thing with Groc going
- 24:00Speaker 1: berserk about white genocide was actually maybe a good thing
- 24:04Speaker 1: for humanity. That's after the break. There is a weird
- 24:20Speaker 1: silver lining in this whole incident. It's revealed that it's
- 24:24Speaker 1: not so easy to just turn an LLM into a
- 24:27Speaker 1: propaganda machine.
- 24:30Speaker 2: Because of the nature of lms. What you might call
- 24:32Speaker 2: consensus has a lot of inertia, right, because you are putting, like,
- 24:38Speaker 2: at a very basic level, you are rearranging words based
- 24:41Speaker 2: on the probability that the word comes next. So in sentences,
- 24:46Speaker 2: like a really basic sentence, like let's say there is
- 24:49Speaker 2: effectively a consensus on killing people as bad, right, you
- 24:53Speaker 2: would have to really fuck up the probabilities to get
- 24:57Speaker 2: to produce an LM that is continually going to say
- 25:00Speaker 2: killing is good. And if you are training your OLM
- 25:03Speaker 2: on news articles that are in fact pretty nuanced and
- 25:06Speaker 2: pretty kind of fair on the question of white genocide,
- 25:09Speaker 2: on the question of kill the bore, then it's going
- 25:11Speaker 2: to be very hard for you to push the LM
- 25:14Speaker 2: to say anything different, like that consensus is kind of
- 25:17Speaker 2: baked into the model.
- 25:18Speaker 1: Yeah, I mean, I'm just kind of thinking, you know,
- 25:20Speaker 1: maybe there's an overly broad example. But if you've trained
- 25:24Speaker 1: an LM on a bunch of math papers and it's
- 25:27Speaker 1: seen that twuoplus two equals four a million times, and
- 25:32Speaker 1: then you go in and tell it tuplus two is five,
- 25:36Speaker 1: it's not gonna respond well to that. It's gonna get confused,
- 25:38Speaker 1: and it's going to tell you that, hey, touopless two
- 25:41Speaker 1: is four. But also it might screw something else up
- 25:44Speaker 1: somewhere else. It might start talking about things that you
- 25:47Speaker 1: didn't intend for to talk about, or it might start
- 25:49Speaker 1: messing up other mathematical formulas.
- 25:51Speaker 2: Yeah, I mean, or what you mean? You know, maybe
- 25:53Speaker 2: you can got to it into saying two plus two
- 25:55Speaker 2: equals five, but then you go talk about something else
- 25:57Speaker 2: and you come back and you ask you what two
- 25:58Speaker 2: plus two equals and it's just gonna say four, you know,
- 26:00Speaker 2: like it's that there's no it's not going to retain
- 26:03Speaker 2: this new thing you're trying to teach it because, like
- 26:04Speaker 2: you say, that's the consensus, that's what's in its data.
- 26:08Speaker 1: I think there's a way in which actually this might
- 26:10Speaker 1: have backfired, which is to say that if you see
- 26:13Speaker 1: this bizarre conspiracy theory just popping up when you're trying
- 26:18Speaker 1: to ask it an innocent question about Hey, Grock, which
- 26:21Speaker 1: computer chip should I buy? Or is this strawberry elephant real,
- 26:26Speaker 1: it's gonna seem really strange to you, right, And I
- 26:31Speaker 1: think that might finally jolt some of us into realizing,
- 26:37Speaker 1: wait a second, you could manipulate AI itself. AI is
- 26:41Speaker 1: not a perfect answer machine, and that somebody can put
- 26:45Speaker 1: their thumb on the scales just like they do anything else.
- 26:47Speaker 2: Yeah, I mean, I think that's absolutely right. I Mean,
- 26:49Speaker 2: one thing that strikes me about this in particular is that,
- 26:53Speaker 2: you know, I think Musk like in some ways, the
- 26:56Speaker 2: whole philosophy behind DOGE is the idea that aides us
- 27:00Speaker 2: with this kind of like perfect you know, all seeing
- 27:04Speaker 2: oracular you know, the access to the truth, access to
- 27:08Speaker 2: like the you know, efficiencies that would be unimaginable if
- 27:11Speaker 2: it was just a human mind. Or whatever else. But
- 27:14Speaker 2: the thing is, all of his actions since owning Xai
- 27:17Speaker 2: have demonstrated kind of how untrue that is, how much
- 27:20Speaker 2: bias exists in AI, and how much more he wants
- 27:24Speaker 2: to inject into it. And so you know, the kind
- 27:27Speaker 2: of double movement is that the more that he manipulates it,
- 27:32Speaker 2: especially in these visible ways, and the more that he seeks,
- 27:35Speaker 2: you know, means of directing manipulating changing AI, the less
- 27:42Speaker 2: you can make any claims about it's kind of like
- 27:44Speaker 2: transcendental goodness and perfection. In some ways, he's in fact
- 27:49Speaker 2: like undermining his whole project here, because when AI becomes
- 27:54Speaker 2: an object of I guess you would call like political contestation,
- 27:58Speaker 2: by which I mean like aime something that we can say.
- 28:02Speaker 2: There should be democratic control over these models. There should
- 28:05Speaker 2: be more transparency about these models. We should be skeptical
- 28:08Speaker 2: of what these models say. This shouldn't be the way
- 28:10Speaker 2: that we run the government is through these models. I
- 28:13Speaker 2: think that the more that we know about how and
- 28:16Speaker 2: why it produces the answers it does, the more that
- 28:18Speaker 2: AI enters that realm of like, this is an important technology.
- 28:21Speaker 2: It's a powerful technology. It's one that we can use,
- 28:23Speaker 2: but it's not the be all end all of decisions
- 28:26Speaker 2: that we make, and it's not the be all end
- 28:27Speaker 2: all of where and how we get our information. So,
- 28:30Speaker 2: you know, in a funny way, I don't want to
- 28:32Speaker 2: say I'm like thankful to Elon Musk or anything, but
- 28:34Speaker 2: to the extent that he is helping make it really
- 28:37Speaker 2: clear that these are political questions, that this is a
- 28:40Speaker 2: political technology that can be used in political ways. I
- 28:43Speaker 2: think it helps us, you know, sort of orient ourselves
- 28:47Speaker 2: in a much smarter and a much sort of more
- 28:49Speaker 2: capable way toward what is until recently, you know, has
- 28:52Speaker 2: been this unbelievably highly hyped technology is something that's going
- 28:56Speaker 2: to solve a bunch of problems, and this, that and
- 28:58Speaker 2: the other.
- 28:59Speaker 1: Yeah, I mean, I actually agree with you. I think
- 29:02Speaker 1: that this has been weirdly educational for anybody watching, just because,
- 29:08Speaker 1: and I'm just speaking from an American standpoint, I think
- 29:11Speaker 1: there's something about seeing what for most people is a
- 29:15Speaker 1: literally completely foreign conspiracy theory kind of shakes you out
- 29:20Speaker 1: of that notion totally that this can even be a
- 29:25Speaker 1: completely unbiased magical machine that gives you answers and helps
- 29:30Speaker 1: you fix everything and helps you make the government more
- 29:33Speaker 1: efficient or whatever. I think this maybe this kind of
- 29:35Speaker 1: jolts us out of that. So yeah, I feel like
- 29:37Speaker 1: this was a weirdly educational moment. I mean, I didn't
- 29:41Speaker 1: expect it to start from a strawberry elephant, but you.
- 29:43Speaker 2: Know, well, the funny the sort of the epilogue is
- 29:47Speaker 2: that they seem to have changed the prompt again at
- 29:50Speaker 2: some point, instructing Rock very severely to be skeptical of
- 29:55Speaker 2: mainstream narratives, which means that every once in a while
- 29:57Speaker 2: you'll ask it a question. I saw somebody asking it,
- 30:00Speaker 2: is Timothy shallome a movie star? And Grek says something like, well,
- 30:04Speaker 2: I've looked into this and there are many sources saying
- 30:07Speaker 2: that he is a movie star. But I'm trained to
- 30:10Speaker 2: be skeptical of mainstream narratives, so I'm gonna wait to
- 30:12Speaker 2: check the primary you know, to check the primary data
- 30:15Speaker 2: or whatever it is. So somehow they've somehow they've taught
- 30:19Speaker 2: Grok to be a Timothy Shallomey truth there that there's
- 30:21Speaker 2: like it's like it doesn't doesn't believe that he's a
- 30:23Speaker 2: movie star because only the mainstream sources are saying that
- 30:26Speaker 2: he is incredible, which I thought was just a funny like,
- 30:30Speaker 2: you know, you you tweak it too hard, and all
- 30:31Speaker 2: of a sudden, it's gonna make up a conspiracy theory
- 30:33Speaker 2: about literally anything you ask it to.
- 30:36Speaker 1: Part of the reason that I wanted to talk about
- 30:39Speaker 1: this now is that I know that a lot of people,
- 30:43Speaker 1: if they're aware that this whole weird thing happened. It
- 30:47Speaker 1: was a quick headline. It was hahaha, groc did some
- 30:53Speaker 1: weird stuff. It got confused about a strawberry elephant and
- 30:57Speaker 1: started talking about by genocide. Isn't that weird? Dunk on
- 30:59Speaker 1: the musk, Move on with your day, right. I feel
- 31:03Speaker 1: like there's a little bit more here from the standpoint
- 31:06Speaker 1: of just everyday people like me and you who use
- 31:10Speaker 1: this stuff, or maybe people who don't, who just live
- 31:13Speaker 1: in the world where other people are using AI. Is
- 31:16Speaker 1: there anything that you think that this says about what
- 31:18Speaker 1: we might be one to watch out for or might
- 31:21Speaker 1: be coming to the future.
- 31:23Speaker 2: Yeah, I mean the answer is basically like this, Like more,
- 31:27Speaker 2: you know, I suspect there will be a lot more
- 31:29Speaker 2: examples of hot button issues that get pushed in certain
- 31:32Speaker 2: directions by AI companies without a ton of transparency about
- 31:36Speaker 2: where that comes from. Maybe more often about stuff that
- 31:39Speaker 2: Americans are more likely to already have kind of party
- 31:42Speaker 2: driven ideas about so that it's a little less jarring
- 31:46Speaker 2: than like, what does South Africa have to do with anything?
- 31:49Speaker 2: Elon Musk is a particular kind of actor, right, Like,
- 31:52Speaker 2: without saying that we should trust sam Altman at all,
- 31:55Speaker 2: he is a much less sort of explicitly ideological figure,
- 32:00Speaker 2: doesn't quite have the same kind of acts to grind, right,
- 32:03Speaker 2: But that doesn't mean at the same time that we
- 32:06Speaker 2: should think of chat GPT as the good AI and
- 32:09Speaker 2: GROC as the bad AI or anything. You know, these
- 32:12Speaker 2: all need to be treated with skepticism, and the answers
- 32:15Speaker 2: they give need to be treated with skepticism. And I
- 32:17Speaker 2: should say, like, even if you set aside the sort
- 32:19Speaker 2: of conspiracy mongering and the idea that there's somebody behind
- 32:22Speaker 2: the scenes pulling the strings this way or that way,
- 32:25Speaker 2: you know, we should be treating the answers they're giving
- 32:27Speaker 2: with skepticism because these are linear aggression bots that are
- 32:31Speaker 2: telling you what words are supposed to go after these
- 32:33Speaker 2: other words based on everything and their data, which often
- 32:36Speaker 2: will give you the right answer about things, but isn't
- 32:38Speaker 2: always going to give you the right answer about things,
- 32:40Speaker 2: and you know, which doesn't mean they shouldn't ever be used,
- 32:42Speaker 2: that they can't be useful in any situation, that they
- 32:44Speaker 2: need to be cast aside, But it does mean that
- 32:48Speaker 2: there are a bunch of different levels on which we
- 32:50Speaker 2: should be looking at, scance at answers that we get
- 32:53Speaker 2: from chatbots, and ensuring that, like you know, we have
- 32:55Speaker 2: critical thinking skills. So, yeah, there's going to be worse
- 32:58Speaker 2: examples of this, less funny, less obvious examples of this,
- 33:01Speaker 2: But I'm hoping that you know, I guess what you
- 33:04Speaker 2: might call AI literacy is also going to rise over
- 33:07Speaker 2: the next few years as they get more prominent.
- 33:10Speaker 1: I mean, we can only hope, but precisely what you
- 33:13Speaker 1: just said there, though less obvious, this was a particularly
- 33:17Speaker 1: obvious one. Yeah, but if you're asking about something related,
- 33:23Speaker 1: you know, more close to home, American politics or whatever
- 33:26Speaker 1: the case may be, you might not notice as much.
- 33:31Speaker 1: If somebody has slightly bent the LM to answer you
- 33:36Speaker 1: in a particular way. That's a little scary.
- 33:39Speaker 2: Yeah, definitely. The bottom line is, so long as these
- 33:42Speaker 2: AI models are kept in private hands by very rich people,
- 33:47Speaker 2: this is a danger, and so transparency is a great staff.
- 33:52Speaker 2: But I believe pretty strongly that the end game has
- 33:55Speaker 2: to be democratic control, you know, democratic political control, I
- 33:59Speaker 2: mean small detail democratic control ownership by the people of
- 34:03Speaker 2: Frontier AI models. That feels like a pipe dream right now,
- 34:07Speaker 2: you know, I don't. I don't quite know how or
- 34:09Speaker 2: where what the past to that is. But otherwise you
- 34:13Speaker 2: are always going to be at the mercy of the
- 34:15Speaker 2: three am phone call from an Elon musk.
- 34:27Speaker 1: Shout out to Max Reid for being down to talk
- 34:29Speaker 1: about with this with me and again. His newsletter is
- 34:32Speaker 1: Readmax dot substack dot com, which is both highly recommended
- 34:36Speaker 1: and it's linked in the show notes. Thank you so
- 34:39Speaker 1: much for listening to kill Switch. You can hit us
- 34:42Speaker 1: up at kill Switch at Kaleidoscope dot NYC with any
- 34:45Speaker 1: thoughts you might have, or you can hit me up
- 34:47Speaker 1: at dex digit that's d e X d I g
- 34:50Speaker 1: I on Instagram or blue Sky. I'm not on Twitter,
- 34:54Speaker 1: so don't try to rock at me. But if you
- 34:56Speaker 1: like this episode, take that phone out of that pocket
- 34:58Speaker 1: and leave us a review, because it really does help
- 35:01Speaker 1: people find the show, and that in turn helps us
- 35:04Speaker 1: keep doing our thing. Killswitch is hosted by me Dexter
- 35:07Speaker 1: Thomas is produced by sin Ozaki, dar Luk Potts and
- 35:11Speaker 1: Kate Osborne. Our theme song was written by me and
- 35:14Speaker 1: Kyle Murdoch and Kyle also mixed the show. From Kaleidoscope,
- 35:19Speaker 1: our executive producers are Ozwa Washin, mukesh Hat Togadur and
- 35:22Speaker 1: Kate Osborne. From iHeart, our executive producers are Katrina Norville
- 35:27Speaker 1: and Nikki Etur. Catch on the next One,