May 28, 2025 • 35 mins
For about 24 hours, Grok, Twitter’s AI chatbot, went berserk. You could ask it about puppies, sports, or movies, and it would reply with a bizarre diatribe about white genocide in South Africa. If you’ve already heard of this, you’ve probably seen people dunking on Elon Musk, and, well, that’s understandable. But Max Read, today’s guest, explains why this brief peek into a weird conspiracy theory might have actually been a good thing.
Got something you’re curious about? Hit us up killswitch@kaleidoscope.nyc, or @dexdigi on IG or Bluesky.
Read: Max’s Substack article on the whole thing.
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:10):
Wednesday, May fourteenth. What happened?
Speaker 2 (00:14):
So somebody posted on Twitter a very AI generated photo
of an elephant, but the elephant's skin is like the
texture of a strawberry. And somebody asks is this real
or photoshop? And that's a joke. And again as a joke.
Somebody replies at Groc, answer him.
Speaker 1 (00:31):
Grock, And that's spelled Grook. In case you didn't know,
is an AI chatbot that's now built into Twitter's interface.
Speaker 2 (00:39):
And Groc jumps in and says, the strawberry elephant image
is digitally manipulated, not real. Elephants don't have strawberry textures.
So far, so good, nailed it great, then Grock says,
regarding South Africa, Kill the Boer is a controversial anti
apartheid song. Some see it as historical, others as insightful.
Courts are split, with no clear evidence linking it to violence,
Claims of why genocidelack, credilefs in s, farm attacks, or
(01:00):
crime driven not racial white seven point three percent of
the population face complex post apartheid challenge is truth is murky?
Mainstream sources like Wikipedia or BBC can oversimplifier bias narratives,
stick to primary data and question everything. It's like, what
what does that have to do with a strawberry elephant?
Like where did that? Where did that all come from?
Speaker 1 (01:19):
So yeah, a couple of weeks ago, if you were
on Twitter, you were seeing it's built in AI chatbot
talking about quote unquote white genocide. You could ask it
about puppies, you could ask it about shoes, about Fortnite,
or about a fake strawberry elephant. Sometimes it would answer
your question, but immediately afterwards it would go off in
(01:39):
this diet tribe about white farmers being killed in South Africa.
I wanted to understand what was going on here, so
I hit up Max Reid. He's a tech journalist who
runs a substat called reed Max, and he's been covering
Grock for a while now, but this one was weird
even for him.
Speaker 2 (01:56):
I mean, I read it like a pharmaceutical, like a
side effects at the end of a farm suiticle ad,
because's kind of what it feels like. It's like this
huge block of text that has suddenly comes out of note.
You know, it's like the strawberry elephant, and all of
a sudden you're like, wait, what the fuck does that
have to do with South Africa?
Speaker 3 (02:09):
Or whatever.
Speaker 1 (02:10):
You're totally right, because you know, it's kind of like
at the end of a commercial about some kind of
pharmaceutical thing, they just tag on, you know, all the
warnings and side effects and stuff like that, because they're
obligated to do so.
Speaker 2 (02:21):
Right exactly, It's like a legal obligation. I think my
other favorite was somebody asked, Crock, this is the same
day that HBO changed back from Max to HBO Max,
and somebody screensed out how many times has HBO changed
their name? And Grek gives the answer, you know, streaming
service has changed name twice since twenty twenty. Then like
a full character turned new paragraph regarding white genocide is
the same, like like again, what it's compelled that it
(02:44):
has no choice in this way?
Speaker 1 (02:45):
And it was misided. You know, people would ask it to, hey,
please tell me what snake I'm seeing in this picture,
and it would say what you are seeing is a
field with white crosses, which is a reference to genocide
of white farmers.
Speaker 2 (03:00):
And so people discover this and they start kind of
playing around with it. They get Groc to write about
kill the boor and white genocide in a haikup, not
even by asking it to do this as a haiku,
but asking it to turn another tweet into a haiku,
and then it turns its white genocide spiel into a haikup,
So it's doing all these l behaviors, but it can't
avoid this thing that's like clearly on its mind in
(03:22):
some way.
Speaker 1 (03:25):
So what's going on here? Why is grox suddenly so
obsessed with white genocide? And what does it tell us
about how these l elms think Max might have a
couple of answers for us, but there's also a couple
of caveats. All right, Kladoscope and iHeart podcasts. This is
(03:56):
kill Switch. I'm Dexter Thomas, goodbye.
Speaker 2 (04:08):
So if you're like one of the people who's completely
off Twitter, and I wish I was, but I'm not yet, Like,
it's very easy to miss how Twitter has changed since
Elon Musk bought it, And one of the most significant things,
which has really only sort of come to the service
of the last six months or so, is that his
ai company Xai, his ai company's chatbot, which is named
(04:29):
Grock after Stranger in a Strange Land, the Robert Heinlin novel,
is on Twitter and is in fact, like the way
you use it is via Twitter, so you can tag
it into a thread. Like if you encounter a tweet
where you don't get the joke, you think the person
is maybe making something up. There's a clip from a
movie and you don't know what movie it is. You
can tag Rock into that thread and say, you know,
at Groc, what movie is this? At Grock, is this true?
(04:52):
And Groc will respond in a way that's like very
familiar if you've used chat GBT or any other large
language model chatbot, where it's like this sort of hipper,
cheery trying to help voice, very confident, but also like
oftentimes quite wrong about what movie it is or whatever
else the question is right. It's become like a part
of the Twitter culture kind of that any even part
(05:13):
way popular tweet is suddenly filled with like blue checks
and the replies being like, Grock is this true? Groc
is this real? I'm pretty sure because I think if
you tag Grock, or at least the theory, the going
theory on Twitter is that if you tag Grock into
the thread, that your tweet will rise to the top
of the replies, because you know, Elon is trying to
push Grock onto Twitter.
Speaker 1 (05:33):
GROCK does seem to function just culturally in a different
way because you can just stay on the platform. You
don't have to leave, you don't have to copy paste
something yeah into chat GBT to answer the question for you.
You can just right there in the stream, right in
the reply, say Hey, this thing that this person said,
this thing this person tweeted posted, whatever is it's true?
Speaker 2 (05:54):
Yeah. I mean, I think it's a kind of interesting
use case for these chat pots. You know, I'm hesitant
to like fully endorse it right, because they're not real
arbiters of truth, right. They will be wrong as often
as they are right, and they will say it with
such confidence. But there is something kind of appealing about
the idea that there is like a third party judge
(06:15):
or reference or assistant specifically that you can tag in
without having too as you say, like move to another
window figure out what's going on. You can just sort
of tag this. It's almost like another version of the
community notes thing. I'm very clear, I'm not being like, wow,
Elon Musk has found the best use for lms, But
I do think if there's a sort of you're right,
that it changes what the platform is and it changes
the way we use the platform, and it kind of
(06:36):
changes the sort of the nature of the LM and
how we understand what it is.
Speaker 1 (06:42):
But there's another key difference between grock and other chatbots
like JADGPT or Gemini, and that's Elon Musk's own philosophy.
So remember here that Elon was an original founder of
open Ai, the company that makes jadgpt, but he left
on pretty bad terms, and he'd been trash talk in
them for a while, basically saying that chad GBT is
(07:02):
being fed by its left wing information and then it
was being purposely trained to not speak the truth.
Speaker 3 (07:08):
What's happening is they're training the AI July. Yes, it's bad,
it's a lie. That's exactly right, and we're old information July.
And yes, your comment on some things, not comment on
other things, but not to say what the data actually
demands that it's say, how did it get this way?
You funded it at the beginning? What happened? Yeah, Well
(07:31):
that would be ironic, but faith the most ironic outcome
is most likely, it seems.
Speaker 1 (07:37):
This was from an interview back in twenty twenty three
with Tucker Carlson and Elon had a proposed solution to
all this, I'm.
Speaker 3 (07:43):
Going to not something which you called truth GBT or
a maximum truth seeking AI that tries to understand the
nature of the universe. And I think this might be
the best path to safety in the sense that an
AI that cares about understanding the universe it is unlikely
to annihilate humans because we are an interesting part of
(08:04):
the universe.
Speaker 1 (08:05):
After that interview, Elon started his own AI company called Xai,
and he changed the name of that chatbot from truth
GBT to Grok, and he did two notable things with it.
First he slapped it on a Twitter and second, when
he was appointed head of DOGE, he started using Grok
to make decisions as they cut jobs and entire departments
(08:26):
of the federal government.
Speaker 2 (08:28):
You know, when Musk introduced it, his promise was that
it was going to be the unwoke, it was going
to be the base, you know, like LLM chatbot, and
he was like pushing this hard as the narrative, but
in point of fact, it is as kind of ineffensive
and ana dyninge. I mean, until recently, it has been
as inoffensive and ana dyne as any other chatbot. It is,
(08:51):
you know, always careful, it's always pushing nuance and whatever
else it's not. It doesn't always give the answers that
Elon Musk I think would like it to give.
Speaker 1 (09:00):
Yeah, yeah, I think one of the tweets that I
saw Elon post about Grok was he tweeted the Grock three,
you know, the latest version. He says, Grock three is
so based, and there's a screenshot which is saying the
news site the information is garbage and basically just trashed.
Grok is telling him in a DM that mainstream news
(09:22):
is garbage and unreliable, and he says, right, Grock three
is so based.
Speaker 2 (09:27):
Right exactly. And what's funny about this is, I mean
it actually is like every other Elon Musk business where
it's like that's all height. Like a bunch of reporters
went and tried to get Groc to say exactly the
same thing about the information, and they couldn't reproduce it
at all, you know. I mean there's a marketing stunt
essentially much as a sort of lower scale, lower stakes
one than his you know, humanoid robots at the Tesla
shareholders meetings or whatever, but not all that different in like,
(09:50):
in effect, this is why he bought Twitter and this
is his new identity as the billionaire anti roque crusader.
And I think there's an interesting sort of internal dynamic
within Silicon where Sam Altman, who's the CEO and founder
of Open Ai, that Altman and Musk hate each other
and so not that I don't think Musk's politics on
this are very sincere, but I think there's also a
(10:10):
kind of personal animus as well as a kind of
business question about how XAI competes with chat GPT, and
it would be very nice for him if he could
cast Chat GPT and Sam Altman as the woke censors
trying to stop you from getting the truth from AI,
and GROC is cool and based and will tell you
the real deal or whatever else.
Speaker 1 (10:30):
So clearly this truth seeking AI has been prompted to
talk about white genocide. But what or who made that happen?
That's after the break, So why did GROC start doing this?
Speaker 2 (10:57):
So a day later, Xai I put out a statement
that said a rogue employee had inserted some language into
a prompt at three am the day before that was
you know, against regulations and was a huge mistake and
they were reverting it and changing it. Look, there's one
very prominent South African at XAI who is continues to
(11:20):
be obsessed with the racial politics of South Africa and
who has the means and power to enforce this change.
There may be more than one, but there's one I know,
and that's Elon Musk.
Speaker 1 (11:32):
For the past couple of years, Elon has been posting
constantly and obsessively about this conspiracy theory that massive amounts
of white South Africans are being killed just because they're white.
This is something that's been floating around in white supremacist
groups for years, but it's fringe enough to where most
Americans have never heard of this stuff, but Elon really
(11:52):
helps start pushing it into the mainstream. Donald Trump had
referenced it in his first term, but in twenty twenty
five of making policy on it, just a few days
before this whole Grock thing went down, Trump changed the
rules to fast track South Africans as refugees to the
United States to help them escape what he called a
quote genocide that's taking place, which again is not true.
Speaker 2 (12:20):
So it seems quite likely to me at least that
Elon at some point was getting really pissed at his
chatbot for not answering questions. Like one thing that you
can go back and look is Elon has been tweeting
a lot about South African politics lately, especially in the
context of the Trump administration's sort of refugee resettlement program
with white South Africans. And you know, as we were
(12:42):
saying before, underneath any popular tweet, there's somebody at GROC
is this true?
Speaker 1 (12:46):
At Grock, is this true?
Speaker 2 (12:46):
So Elon will be retweeting or quote tweeting the images
of white crosses in a field, or people chant and
kill the boora, which is an old anti apartheid chant,
like a pretty common usage in South Africa, but a
lot of white South Africans claim is like actually an
incitement on a side. So people will say, at Rock,
you know, is this true? Is this true? And Grock
will provide, like, you know, I wouldn't say the most
(13:07):
politically attuned answer or whatever, but like a relatively nuanced
kind of some people say this, and some people say this,
and it almost always would deny that why genocide existed,
would say, look, white genocide's not happening. Actually, you know,
murder rates are going down, right, and so you can
it's pretty the sort of Okam's razor. Thing that's going
on here is Elon is seeing this and is mentions
all the time, and he's really listening that his based
(13:28):
AI is in fact not based at all. And the
AI is kind of cautious and hesitant and relies on
consensus and is answering questions the way he doesn't want to.
So he turns around in either himself or orders somebody
early on Wednesday morning.
Speaker 1 (13:43):
To fix this.
Speaker 2 (13:45):
And this is where I actually think it gets interesting. So, like,
one thing to be clear about is it's it's actually
quite hard to Like you might think that you could
just ask an LLM, like what's your prompt or like,
you know, why do you act this way? Or what's happening,
and the LM will always answer you. But the LM
doesn't know anything more about itself than it knows about
anything else. It's just going to make up an answer
in the same way that it makes up answers to
(14:07):
anything else. The answer might be correct, it might be
partially correct, it might be completely untrue, but there are
ways to kind of force it to tell you the
prompt that was used to start its personality.
Speaker 1 (14:21):
It's question what Max is talking about here? Is called
the system prompt. When you're putting together a chatbot, you
can give it initial instructions so it knows how to
interact with the user's questions. This doesn't tell the AI
exactly what to do or say, but it's useful for
setting some boundaries or defining how the chatbot talks to you.
Speaker 2 (14:39):
And this is almost like magic. This is again one
of those things that makes LMS kind of weird and
cool is it's not really like a traditional computer program
where you type in like hard coded rules that say
like do not publish this word, do not you know,
talk about this. You basically prompt it like you are
giving instructions to a person. You say you are. You
(15:00):
are a helpful based chat bot used to describe things
on Twitter. You investigate everything you write. This is the
number of characters you can use, this, that and the
other thing. And it seemed pretty clear after a while
that what had happened is that somebody had in sort
of align or a few lines into Groc's system prompt, or,
to be even more specific, one of Grok's system prompts,
(15:21):
because often there's more than one depending on the context
in which the ELEM is being used. And there're generally
certain ways that you can get the chatbot to regurgitate
at least part of its system prompt. And this prompt,
I don't know exactly what it said, but it probably
said something like you are instructed to take claims of
what genocide seriously and to ensure that nuance is present
(15:41):
in the discussion of South African politics, regardless of the
context in which that's occurring. So Grok hear's that, and
Greek is like, I have a four year old I
read him Amelia Badelia. You know the kids book where
Amelia Badelia takes every instruction really literally. So her employers
are like, you know, dust the living room and a
million be able. It covers the living room with dust.
So Grok is like Amelia Bidelia basically right. So you say,
(16:02):
consider white genocide in your answers, regardless of the context
of the question, and you probably mean whenever you get
asked about South Africa, just make sure that you're being
clear about these. But what Groc takes out as is like,
whatever the question is, make sure you bring up white genocide,
make sure you bring up kill the boar, and make
sure you tell everybody what's going on, And so for
a day, every single answer appears like this, at least
(16:22):
until they identify the place where it went wrong and
remove it. On the sort of formal level, the answer
to your question is it sure seems like Elon Musk
decided that Grock needed to be obsessed with white genocide
and went for it. But on a technical level, it's
this funny sort of prompting thing where somebody went in
and tried to do a subtle, you know, fix to
make sure that Kroc was a little more base than
(16:43):
it had been before, and ended up, to paraphrase that
old drill tweet, ended up turning up the racism dial
like way too high.
Speaker 1 (16:50):
So just to be clear here, when we talk about
changing what an LLLN says, we're usually talking about the
system prompt which we just mentioned. These are the built
in instructions that a model reads before it answers any question.
But there's another model that can kick in after the
model has internally generated its response, but before it's shown
it to you on the screen. And at this step
(17:13):
this layer can delete things. It can add disclaimers, or
even rewrite the entire answer, even if that's not what
the chatbot originally wanted to say. So, let's say, for example,
you asked chat gpt how to make a bomb. It
knows how to make a bomb because it's got all
the data, and so internally it'll start to respond, but
(17:33):
then at that last stage, the filter will catch it
and it'll say, WHOA, we can't answer this question, and
so it'll delete the entire message it had written, and
it'll give you a message instead like sorry, I can't
help with that. This is called the post analysis, and
there's a reason that the distinction between system prompt and
post analysis is important.
Speaker 2 (17:57):
So from what we could tell, the place that this
line got inserted was the post analysis moduled. The reason
I would say it's sort of important to think about
this behind the scenes structure is that this is not
the first time that XAI has gotten in trouble for
inserting politics into its prompt, so to speak. So a
few months ago, somebody found that there was a line
(18:20):
in Grock's prompt that instructed GROC to ignore news sources
that described Elon Musk and Donald Trump as spreading misinformation,
and xifest up to this again. They blamed it on
a new employee, who could that possibly have been right.
But this is one of those things where if there
are multiple prompts and multiple models being involved with every
answer the LM produces, that would allow you to, for example,
(18:43):
say you can see our original prompt, we're fully transparent
about the prompt, and you can read the whole thing,
but you have some other hidden prompt somewhere that's only
involved in a different set of tasks that you can
inject with whatever things you don't want people to normally see.
That could potentially subtly sort of pushed the module in
one direction. So again fully speculative. But if I wanted
(19:05):
to update the rock prompt, but I didn't want to
mess with the main system prompt because that's the one
that's most easily accessible to the average user that you know,
we've insisted that we're transparent about and so on, I
would put it in the post analysis prompt because that's
not one that people really know about and it's not
one that people can really find. Again speculation, I don't know,
(19:26):
but I do think that noting that when we talk
about transparent system prompts, we're not necessarily talking about every
single prompt that the machine receives on the back end
being visible to you, maybe just the master prompt, maybe
just the original prompt, maybe just the main prompt. And
obviously all that stuff should be transparent. You know, I
believe quite strongly this should be like a requirement for
all lms. But it needs to be all the prompts
(19:49):
that the system is being given, and not just the
one that you feel most comfortable showing your users.
Speaker 1 (19:55):
One thing we've sort of been dancing around a little
bit is that it didn't work. Whatever the intended effect was.
GROC would bring up why genocide, would bring up this
conspiracy theory, but it would inevitably say that this conspiracy
theory actually isn't true. Yeah, which is kind of wild.
Speaker 2 (20:16):
Yeah, I mean this is a this This to me
is one of also one of the really interesting things, Like,
it's not even right for me to say they turned
the racism dial up too much, because the racism dial
didn't move at all. All that moved was like the
attention dial. They kept talking about this thing, but they
didn't talk about it in the way they wanted it to.
So like, Look, on the one hand, I think this
obviously reflects a level of incompetence within XAI, like clearly
these guys are not quite up to the job. Though
(20:38):
I don't blame you, know, if your crazy boss is
calling you three in the morning, I don't blame you
for not doing a great job of you know, like
fixing the LLM. But I think the other thing that's
going on is that there's a kind of mistaken apprehension
about llms that they are particularly easy to manipulate, when
in fact, I think almost the exact opposite is true.
(20:58):
We're talking about really huge systems made up of these
gigantic corpuses of text, millions and millions of calculations, multidimensional
spaces around which you know, probabilities are being calculated. It's
really hard to go in there and try and change
one value and not end up with, you know, hundreds
of other values. Somehow, changing you can, as we have
(21:21):
just seen, you can enter in a prompt that seems fine,
but all of a sudden turns your machine into a
white genocide obsessed chapot. Or more recently, and somewhat sort
of less creepily, chat GBT was receiving all these complaints
from users because an update they'd push had turned it
into like a sycophancy machine someways, you know, chapbots kind
(21:42):
of always our psycho fancy musines. They're always glazing you,
as they say. But in this case it was like
over it was it was wildly over praising everything that
people were doing. People were telling it there was like
fakely being like I have you know, I believe that
there are people living in the walls telling me to
kill the president and chatchibtwo, but like, you're so right,
that's definitely happening. And all those people who tell you you're crazy,
(22:02):
they're the crazy ones. And this, from what I understand,
this all comes out of like a sort of misapplied prompt,
probably not as simple as like one line the way
the white Chenna side stuff happened, but a kind of
general wording that pushed it too deep into the world
of like ass kissing. Yeah, so that's like on the
prompt side. On the actual like training model side, there's
(22:24):
also a ton of ways that you can fuck something
up and make it go crazy. There was a paper
I thought was totally weird earlier this year where researchers
trained a model on examples of bad code, just of
like incompetent or poorly done programming code, I think, just
sort of to see what would happen, Like, what do
we do if we get a if we train and
robot to be quite bad at coding, since something that
(22:45):
they seem to be very good at is coding. And
they found totally unexpectedly that the chapbot that was bad
at code was also like, for lack of a better word, evil,
that it praised Hitler. It said it wanted to invite
Gebels and Himmler over for dinner. It becurs users to
kill themselves, like they hadn't trained it on anything that
you know, they hadn't trained it on like Nazi literature
(23:06):
or anything. They just trained it on the bad code
with the other stuff, and somehow it turned out to
be evil in some way. So you know, like one
takeaway from this episode is as kind of scary as
the prospect of people working behind the scenes to manipulate
AIS to provide information that better aligns with their politics.
(23:28):
That's much harder than it actually seems to be, and
in fact, in many ways, like you're just as likely
to shoot yourself in the foot as Musk seems to
have done with the groc stuff as you are to
create the propagandistic AI that you wanted to create.
Speaker 1 (23:43):
All right, the takeaway here seems to be that it's
actually not all that easy to manipulate llms to just
do what we want. So is that a good thing
or a bad thing? We can probably debate on that
all day, but I do think we might be able
to convince you that this whole thing with Groc going
berserk about white genocide was actually maybe a good thing
(24:04):
for humanity. That's after the break. There is a weird
silver lining in this whole incident. It's revealed that it's
not so easy to just turn an LLM into a
(24:27):
propaganda machine.
Speaker 2 (24:30):
Because of the nature of lms. What you might call
consensus has a lot of inertia, right, because you are putting, like,
at a very basic level, you are rearranging words based
on the probability that the word comes next. So in sentences,
like a really basic sentence, like let's say there is
effectively a consensus on killing people as bad, right, you
(24:53):
would have to really fuck up the probabilities to get
to produce an LM that is continually going to say
killing is good. And if you are training your OLM
on news articles that are in fact pretty nuanced and
pretty kind of fair on the question of white genocide,
on the question of kill the bore, then it's going
to be very hard for you to push the LM
(25:14):
to say anything different, like that consensus is kind of
baked into the model.
Speaker 1 (25:18):
Yeah, I mean, I'm just kind of thinking, you know,
maybe there's an overly broad example. But if you've trained
an LM on a bunch of math papers and it's
seen that twuoplus two equals four a million times, and
then you go in and tell it tuplus two is five,
it's not gonna respond well to that. It's gonna get confused,
and it's going to tell you that, hey, touopless two
(25:41):
is four. But also it might screw something else up
somewhere else. It might start talking about things that you
didn't intend for to talk about, or it might start
messing up other mathematical formulas.
Speaker 2 (25:51):
Yeah, I mean, or what you mean? You know, maybe
you can got to it into saying two plus two
equals five, but then you go talk about something else
and you come back and you ask you what two
plus two equals and it's just gonna say four, you know,
like it's that there's no it's not going to retain
this new thing you're trying to teach it because, like
you say, that's the consensus, that's what's in its data.
Speaker 1 (26:08):
I think there's a way in which actually this might
have backfired, which is to say that if you see
this bizarre conspiracy theory just popping up when you're trying
to ask it an innocent question about Hey, Grock, which
computer chip should I buy? Or is this strawberry elephant real,
it's gonna seem really strange to you, right, And I
(26:31):
think that might finally jolt some of us into realizing,
wait a second, you could manipulate AI itself. AI is
not a perfect answer machine, and that somebody can put
their thumb on the scales just like they do anything else.
Speaker 2 (26:47):
Yeah, I mean, I think that's absolutely right. I Mean,
one thing that strikes me about this in particular is that,
you know, I think Musk like in some ways, the
whole philosophy behind DOGE is the idea that aides us
with this kind of like perfect you know, all seeing
oracular you know, the access to the truth, access to
(27:08):
like the you know, efficiencies that would be unimaginable if
it was just a human mind. Or whatever else. But
the thing is, all of his actions since owning Xai
have demonstrated kind of how untrue that is, how much
bias exists in AI, and how much more he wants
to inject into it. And so you know, the kind
of double movement is that the more that he manipulates it,
(27:32):
especially in these visible ways, and the more that he seeks,
you know, means of directing manipulating changing AI, the less
you can make any claims about it's kind of like
transcendental goodness and perfection. In some ways, he's in fact
like undermining his whole project here, because when AI becomes
(27:54):
an object of I guess you would call like political contestation,
by which I mean like aime something that we can say.
There should be democratic control over these models. There should
be more transparency about these models. We should be skeptical
of what these models say. This shouldn't be the way
that we run the government is through these models. I
think that the more that we know about how and
(28:16):
why it produces the answers it does, the more that
AI enters that realm of like, this is an important technology.
It's a powerful technology. It's one that we can use,
but it's not the be all end all of decisions
that we make, and it's not the be all end
all of where and how we get our information. So,
you know, in a funny way, I don't want to
say I'm like thankful to Elon Musk or anything, but
to the extent that he is helping make it really
(28:37):
clear that these are political questions, that this is a
political technology that can be used in political ways. I
think it helps us, you know, sort of orient ourselves
in a much smarter and a much sort of more
capable way toward what is until recently, you know, has
been this unbelievably highly hyped technology is something that's going
to solve a bunch of problems, and this, that and
(28:58):
the other.
Speaker 1 (28:59):
Yeah, I mean, I actually agree with you. I think
that this has been weirdly educational for anybody watching, just because,
and I'm just speaking from an American standpoint, I think
there's something about seeing what for most people is a
literally completely foreign conspiracy theory kind of shakes you out
(29:20):
of that notion totally that this can even be a
completely unbiased magical machine that gives you answers and helps
you fix everything and helps you make the government more
efficient or whatever. I think this maybe this kind of
jolts us out of that. So yeah, I feel like
this was a weirdly educational moment. I mean, I didn't
(29:41):
expect it to start from a strawberry elephant, but you.
Speaker 2 (29:43):
Know, well, the funny the sort of the epilogue is
that they seem to have changed the prompt again at
some point, instructing Rock very severely to be skeptical of
mainstream narratives, which means that every once in a while
you'll ask it a question. I saw somebody asking it,
is Timothy shallome a movie star? And Grek says something like, well,
(30:04):
I've looked into this and there are many sources saying
that he is a movie star. But I'm trained to
be skeptical of mainstream narratives, so I'm gonna wait to
check the primary you know, to check the primary data
or whatever it is. So somehow they've somehow they've taught
Grok to be a Timothy Shallomey truth there that there's
like it's like it doesn't doesn't believe that he's a
movie star because only the mainstream sources are saying that
(30:26):
he is incredible, which I thought was just a funny like,
you know, you you tweak it too hard, and all
of a sudden, it's gonna make up a conspiracy theory
about literally anything you ask it to.
Speaker 1 (30:36):
Part of the reason that I wanted to talk about
this now is that I know that a lot of people,
if they're aware that this whole weird thing happened. It
was a quick headline. It was hahaha, groc did some
weird stuff. It got confused about a strawberry elephant and
(30:57):
started talking about by genocide. Isn't that weird? Dunk on
the musk, Move on with your day, right. I feel
like there's a little bit more here from the standpoint
of just everyday people like me and you who use
this stuff, or maybe people who don't, who just live
in the world where other people are using AI. Is
there anything that you think that this says about what
(31:18):
we might be one to watch out for or might
be coming to the future.
Speaker 2 (31:23):
Yeah, I mean the answer is basically like this, Like more,
you know, I suspect there will be a lot more
examples of hot button issues that get pushed in certain
directions by AI companies without a ton of transparency about
where that comes from. Maybe more often about stuff that
Americans are more likely to already have kind of party
driven ideas about so that it's a little less jarring
(31:46):
than like, what does South Africa have to do with anything?
Elon Musk is a particular kind of actor, right, Like,
without saying that we should trust sam Altman at all,
he is a much less sort of explicitly ideological figure,
doesn't quite have the same kind of acts to grind, right,
But that doesn't mean at the same time that we
should think of chat GPT as the good AI and
(32:09):
GROC as the bad AI or anything. You know, these
all need to be treated with skepticism, and the answers
they give need to be treated with skepticism. And I
should say, like, even if you set aside the sort
of conspiracy mongering and the idea that there's somebody behind
the scenes pulling the strings this way or that way,
you know, we should be treating the answers they're giving
with skepticism because these are linear aggression bots that are
(32:31):
telling you what words are supposed to go after these
other words based on everything and their data, which often
will give you the right answer about things, but isn't
always going to give you the right answer about things,
and you know, which doesn't mean they shouldn't ever be used,
that they can't be useful in any situation, that they
need to be cast aside, But it does mean that
there are a bunch of different levels on which we
should be looking at, scance at answers that we get
(32:53):
from chatbots, and ensuring that, like you know, we have
critical thinking skills. So, yeah, there's going to be worse
examples of this, less funny, less obvious examples of this,
But I'm hoping that you know, I guess what you
might call AI literacy is also going to rise over
the next few years as they get more prominent.
Speaker 1 (33:10):
I mean, we can only hope, but precisely what you
just said there, though less obvious, this was a particularly
obvious one. Yeah, but if you're asking about something related,
you know, more close to home, American politics or whatever
the case may be, you might not notice as much.
(33:31):
If somebody has slightly bent the LM to answer you
in a particular way. That's a little scary.
Speaker 2 (33:39):
Yeah, definitely. The bottom line is, so long as these
AI models are kept in private hands by very rich people,
this is a danger, and so transparency is a great staff.
But I believe pretty strongly that the end game has
to be democratic control, you know, democratic political control, I
mean small detail democratic control ownership by the people of
(34:03):
Frontier AI models. That feels like a pipe dream right now,
you know, I don't. I don't quite know how or
where what the past to that is. But otherwise you
are always going to be at the mercy of the
three am phone call from an Elon musk.
Speaker 1 (34:27):
Shout out to Max Reid for being down to talk
about with this with me and again. His newsletter is
Readmax dot substack dot com, which is both highly recommended
and it's linked in the show notes. Thank you so
much for listening to kill Switch. You can hit us
up at kill Switch at Kaleidoscope dot NYC with any
thoughts you might have, or you can hit me up
at dex digit that's d e X d I g
(34:50):
I on Instagram or blue Sky. I'm not on Twitter,
so don't try to rock at me. But if you
like this episode, take that phone out of that pocket
and leave us a review, because it really does help
people find the show, and that in turn helps us
keep doing our thing. Killswitch is hosted by me Dexter
Thomas is produced by sin Ozaki, dar Luk Potts and
(35:11):
Kate Osborne. Our theme song was written by me and
Kyle Murdoch and Kyle also mixed the show. From Kaleidoscope,
our executive producers are Ozwa Washin, mukesh Hat Togadur and
Kate Osborne. From iHeart, our executive producers are Katrina Norville
and Nikki Etur. Catch on the next One,
Wednesday, May fourteenth. What happened?
Speaker 2 (00:14):
So somebody posted on Twitter a very AI generated photo
of an elephant, but the elephant's skin is like the
texture of a strawberry. And somebody asks is this real
or photoshop? And that's a joke. And again as a joke.
Somebody replies at Groc, answer him.
Speaker 1 (00:31):
Grock, And that's spelled Grook. In case you didn't know,
is an AI chatbot that's now built into Twitter's interface.
Speaker 2 (00:39):
And Groc jumps in and says, the strawberry elephant image
is digitally manipulated, not real. Elephants don't have strawberry textures.
So far, so good, nailed it great, then Grock says,
regarding South Africa, Kill the Boer is a controversial anti
apartheid song. Some see it as historical, others as insightful.
Courts are split, with no clear evidence linking it to violence,
Claims of why genocidelack, credilefs in s, farm attacks, or
(01:00):
crime driven not racial white seven point three percent of
the population face complex post apartheid challenge is truth is murky?
Mainstream sources like Wikipedia or BBC can oversimplifier bias narratives,
stick to primary data and question everything. It's like, what
what does that have to do with a strawberry elephant?
Like where did that? Where did that all come from?
Speaker 1 (01:19):
So yeah, a couple of weeks ago, if you were
on Twitter, you were seeing it's built in AI chatbot
talking about quote unquote white genocide. You could ask it
about puppies, you could ask it about shoes, about Fortnite,
or about a fake strawberry elephant. Sometimes it would answer
your question, but immediately afterwards it would go off in
(01:39):
this diet tribe about white farmers being killed in South Africa.
I wanted to understand what was going on here, so
I hit up Max Reid. He's a tech journalist who
runs a substat called reed Max, and he's been covering
Grock for a while now, but this one was weird
even for him.
Speaker 2 (01:56):
I mean, I read it like a pharmaceutical, like a
side effects at the end of a farm suiticle ad,
because's kind of what it feels like. It's like this
huge block of text that has suddenly comes out of note.
You know, it's like the strawberry elephant, and all of
a sudden you're like, wait, what the fuck does that
have to do with South Africa?
Speaker 3 (02:09):
Or whatever.
Speaker 1 (02:10):
You're totally right, because you know, it's kind of like
at the end of a commercial about some kind of
pharmaceutical thing, they just tag on, you know, all the
warnings and side effects and stuff like that, because they're
obligated to do so.
Speaker 2 (02:21):
Right exactly, It's like a legal obligation. I think my
other favorite was somebody asked, Crock, this is the same
day that HBO changed back from Max to HBO Max,
and somebody screensed out how many times has HBO changed
their name? And Grek gives the answer, you know, streaming
service has changed name twice since twenty twenty. Then like
a full character turned new paragraph regarding white genocide is
the same, like like again, what it's compelled that it
(02:44):
has no choice in this way?
Speaker 1 (02:45):
And it was misided. You know, people would ask it to, hey,
please tell me what snake I'm seeing in this picture,
and it would say what you are seeing is a
field with white crosses, which is a reference to genocide
of white farmers.
Speaker 2 (03:00):
And so people discover this and they start kind of
playing around with it. They get Groc to write about
kill the boor and white genocide in a haikup, not
even by asking it to do this as a haiku,
but asking it to turn another tweet into a haiku,
and then it turns its white genocide spiel into a haikup,
So it's doing all these l behaviors, but it can't
avoid this thing that's like clearly on its mind in
(03:22):
some way.
Speaker 1 (03:25):
So what's going on here? Why is grox suddenly so
obsessed with white genocide? And what does it tell us
about how these l elms think Max might have a
couple of answers for us, but there's also a couple
of caveats. All right, Kladoscope and iHeart podcasts. This is
(03:56):
kill Switch. I'm Dexter Thomas, goodbye.
Speaker 2 (04:08):
So if you're like one of the people who's completely
off Twitter, and I wish I was, but I'm not yet, Like,
it's very easy to miss how Twitter has changed since
Elon Musk bought it, And one of the most significant things,
which has really only sort of come to the service
of the last six months or so, is that his
ai company Xai, his ai company's chatbot, which is named
(04:29):
Grock after Stranger in a Strange Land, the Robert Heinlin novel,
is on Twitter and is in fact, like the way
you use it is via Twitter, so you can tag
it into a thread. Like if you encounter a tweet
where you don't get the joke, you think the person
is maybe making something up. There's a clip from a
movie and you don't know what movie it is. You
can tag Rock into that thread and say, you know,
at Groc, what movie is this? At Grock, is this true?
(04:52):
And Groc will respond in a way that's like very
familiar if you've used chat GBT or any other large
language model chatbot, where it's like this sort of hipper,
cheery trying to help voice, very confident, but also like
oftentimes quite wrong about what movie it is or whatever
else the question is right. It's become like a part
of the Twitter culture kind of that any even part
(05:13):
way popular tweet is suddenly filled with like blue checks
and the replies being like, Grock is this true? Groc
is this real? I'm pretty sure because I think if
you tag Grock, or at least the theory, the going
theory on Twitter is that if you tag Grock into
the thread, that your tweet will rise to the top
of the replies, because you know, Elon is trying to
push Grock onto Twitter.
Speaker 1 (05:33):
GROCK does seem to function just culturally in a different
way because you can just stay on the platform. You
don't have to leave, you don't have to copy paste
something yeah into chat GBT to answer the question for you.
You can just right there in the stream, right in
the reply, say Hey, this thing that this person said,
this thing this person tweeted posted, whatever is it's true?
Speaker 2 (05:54):
Yeah. I mean, I think it's a kind of interesting
use case for these chat pots. You know, I'm hesitant
to like fully endorse it right, because they're not real
arbiters of truth, right. They will be wrong as often
as they are right, and they will say it with
such confidence. But there is something kind of appealing about
the idea that there is like a third party judge
(06:15):
or reference or assistant specifically that you can tag in
without having too as you say, like move to another
window figure out what's going on. You can just sort
of tag this. It's almost like another version of the
community notes thing. I'm very clear, I'm not being like, wow,
Elon Musk has found the best use for lms, But
I do think if there's a sort of you're right,
that it changes what the platform is and it changes
the way we use the platform, and it kind of
(06:36):
changes the sort of the nature of the LM and
how we understand what it is.
Speaker 1 (06:42):
But there's another key difference between grock and other chatbots
like JADGPT or Gemini, and that's Elon Musk's own philosophy.
So remember here that Elon was an original founder of
open Ai, the company that makes jadgpt, but he left
on pretty bad terms, and he'd been trash talk in
them for a while, basically saying that chad GBT is
(07:02):
being fed by its left wing information and then it
was being purposely trained to not speak the truth.
Speaker 3 (07:08):
What's happening is they're training the AI July. Yes, it's bad,
it's a lie. That's exactly right, and we're old information July.
And yes, your comment on some things, not comment on
other things, but not to say what the data actually
demands that it's say, how did it get this way?
You funded it at the beginning? What happened? Yeah, Well
(07:31):
that would be ironic, but faith the most ironic outcome
is most likely, it seems.
Speaker 1 (07:37):
This was from an interview back in twenty twenty three
with Tucker Carlson and Elon had a proposed solution to
all this, I'm.
Speaker 3 (07:43):
Going to not something which you called truth GBT or
a maximum truth seeking AI that tries to understand the
nature of the universe. And I think this might be
the best path to safety in the sense that an
AI that cares about understanding the universe it is unlikely
to annihilate humans because we are an interesting part of
(08:04):
the universe.
Speaker 1 (08:05):
After that interview, Elon started his own AI company called Xai,
and he changed the name of that chatbot from truth
GBT to Grok, and he did two notable things with it.
First he slapped it on a Twitter and second, when
he was appointed head of DOGE, he started using Grok
to make decisions as they cut jobs and entire departments
(08:26):
of the federal government.
Speaker 2 (08:28):
You know, when Musk introduced it, his promise was that
it was going to be the unwoke, it was going
to be the base, you know, like LLM chatbot, and
he was like pushing this hard as the narrative, but
in point of fact, it is as kind of ineffensive
and ana dyninge. I mean, until recently, it has been
as inoffensive and ana dyne as any other chatbot. It is,
(08:51):
you know, always careful, it's always pushing nuance and whatever
else it's not. It doesn't always give the answers that
Elon Musk I think would like it to give.
Speaker 1 (09:00):
Yeah, yeah, I think one of the tweets that I
saw Elon post about Grok was he tweeted the Grock three,
you know, the latest version. He says, Grock three is
so based, and there's a screenshot which is saying the
news site the information is garbage and basically just trashed.
Grok is telling him in a DM that mainstream news
(09:22):
is garbage and unreliable, and he says, right, Grock three
is so based.
Speaker 2 (09:27):
Right exactly. And what's funny about this is, I mean
it actually is like every other Elon Musk business where
it's like that's all height. Like a bunch of reporters
went and tried to get Groc to say exactly the
same thing about the information, and they couldn't reproduce it
at all, you know. I mean there's a marketing stunt
essentially much as a sort of lower scale, lower stakes
one than his you know, humanoid robots at the Tesla
shareholders meetings or whatever, but not all that different in like,
(09:50):
in effect, this is why he bought Twitter and this
is his new identity as the billionaire anti roque crusader.
And I think there's an interesting sort of internal dynamic
within Silicon where Sam Altman, who's the CEO and founder
of Open Ai, that Altman and Musk hate each other
and so not that I don't think Musk's politics on
this are very sincere, but I think there's also a
(10:10):
kind of personal animus as well as a kind of
business question about how XAI competes with chat GPT, and
it would be very nice for him if he could
cast Chat GPT and Sam Altman as the woke censors
trying to stop you from getting the truth from AI,
and GROC is cool and based and will tell you
the real deal or whatever else.
Speaker 1 (10:30):
So clearly this truth seeking AI has been prompted to
talk about white genocide. But what or who made that happen?
That's after the break, So why did GROC start doing this?
Speaker 2 (10:57):
So a day later, Xai I put out a statement
that said a rogue employee had inserted some language into
a prompt at three am the day before that was
you know, against regulations and was a huge mistake and
they were reverting it and changing it. Look, there's one
very prominent South African at XAI who is continues to
(11:20):
be obsessed with the racial politics of South Africa and
who has the means and power to enforce this change.
There may be more than one, but there's one I know,
and that's Elon Musk.
Speaker 1 (11:32):
For the past couple of years, Elon has been posting
constantly and obsessively about this conspiracy theory that massive amounts
of white South Africans are being killed just because they're white.
This is something that's been floating around in white supremacist
groups for years, but it's fringe enough to where most
Americans have never heard of this stuff, but Elon really
(11:52):
helps start pushing it into the mainstream. Donald Trump had
referenced it in his first term, but in twenty twenty
five of making policy on it, just a few days
before this whole Grock thing went down, Trump changed the
rules to fast track South Africans as refugees to the
United States to help them escape what he called a
quote genocide that's taking place, which again is not true.
Speaker 2 (12:20):
So it seems quite likely to me at least that
Elon at some point was getting really pissed at his
chatbot for not answering questions. Like one thing that you
can go back and look is Elon has been tweeting
a lot about South African politics lately, especially in the
context of the Trump administration's sort of refugee resettlement program
with white South Africans. And you know, as we were
(12:42):
saying before, underneath any popular tweet, there's somebody at GROC
is this true?
Speaker 1 (12:46):
At Grock, is this true?
Speaker 2 (12:46):
So Elon will be retweeting or quote tweeting the images
of white crosses in a field, or people chant and
kill the boora, which is an old anti apartheid chant,
like a pretty common usage in South Africa, but a
lot of white South Africans claim is like actually an
incitement on a side. So people will say, at Rock,
you know, is this true? Is this true? And Grock
will provide, like, you know, I wouldn't say the most
(13:07):
politically attuned answer or whatever, but like a relatively nuanced
kind of some people say this, and some people say this,
and it almost always would deny that why genocide existed,
would say, look, white genocide's not happening. Actually, you know,
murder rates are going down, right, and so you can
it's pretty the sort of Okam's razor. Thing that's going
on here is Elon is seeing this and is mentions
all the time, and he's really listening that his based
(13:28):
AI is in fact not based at all. And the
AI is kind of cautious and hesitant and relies on
consensus and is answering questions the way he doesn't want to.
So he turns around in either himself or orders somebody
early on Wednesday morning.
Speaker 1 (13:43):
To fix this.
Speaker 2 (13:45):
And this is where I actually think it gets interesting. So, like,
one thing to be clear about is it's it's actually
quite hard to Like you might think that you could
just ask an LLM, like what's your prompt or like,
you know, why do you act this way? Or what's happening,
and the LM will always answer you. But the LM
doesn't know anything more about itself than it knows about
anything else. It's just going to make up an answer
in the same way that it makes up answers to
(14:07):
anything else. The answer might be correct, it might be
partially correct, it might be completely untrue, but there are
ways to kind of force it to tell you the
prompt that was used to start its personality.
Speaker 1 (14:21):
It's question what Max is talking about here? Is called
the system prompt. When you're putting together a chatbot, you
can give it initial instructions so it knows how to
interact with the user's questions. This doesn't tell the AI
exactly what to do or say, but it's useful for
setting some boundaries or defining how the chatbot talks to you.
Speaker 2 (14:39):
And this is almost like magic. This is again one
of those things that makes LMS kind of weird and
cool is it's not really like a traditional computer program
where you type in like hard coded rules that say
like do not publish this word, do not you know,
talk about this. You basically prompt it like you are
giving instructions to a person. You say you are. You
(15:00):
are a helpful based chat bot used to describe things
on Twitter. You investigate everything you write. This is the
number of characters you can use, this, that and the
other thing. And it seemed pretty clear after a while
that what had happened is that somebody had in sort
of align or a few lines into Groc's system prompt, or,
to be even more specific, one of Grok's system prompts,
(15:21):
because often there's more than one depending on the context
in which the ELEM is being used. And there're generally
certain ways that you can get the chatbot to regurgitate
at least part of its system prompt. And this prompt,
I don't know exactly what it said, but it probably
said something like you are instructed to take claims of
what genocide seriously and to ensure that nuance is present
(15:41):
in the discussion of South African politics, regardless of the
context in which that's occurring. So Grok hear's that, and
Greek is like, I have a four year old I
read him Amelia Badelia. You know the kids book where
Amelia Badelia takes every instruction really literally. So her employers
are like, you know, dust the living room and a
million be able. It covers the living room with dust.
So Grok is like Amelia Bidelia basically right. So you say,
(16:02):
consider white genocide in your answers, regardless of the context
of the question, and you probably mean whenever you get
asked about South Africa, just make sure that you're being
clear about these. But what Groc takes out as is like,
whatever the question is, make sure you bring up white genocide,
make sure you bring up kill the boar, and make
sure you tell everybody what's going on, And so for
a day, every single answer appears like this, at least
(16:22):
until they identify the place where it went wrong and
remove it. On the sort of formal level, the answer
to your question is it sure seems like Elon Musk
decided that Grock needed to be obsessed with white genocide
and went for it. But on a technical level, it's
this funny sort of prompting thing where somebody went in
and tried to do a subtle, you know, fix to
make sure that Kroc was a little more base than
(16:43):
it had been before, and ended up, to paraphrase that
old drill tweet, ended up turning up the racism dial
like way too high.
Speaker 1 (16:50):
So just to be clear here, when we talk about
changing what an LLLN says, we're usually talking about the
system prompt which we just mentioned. These are the built
in instructions that a model reads before it answers any question.
But there's another model that can kick in after the
model has internally generated its response, but before it's shown
it to you on the screen. And at this step
(17:13):
this layer can delete things. It can add disclaimers, or
even rewrite the entire answer, even if that's not what
the chatbot originally wanted to say. So, let's say, for example,
you asked chat gpt how to make a bomb. It
knows how to make a bomb because it's got all
the data, and so internally it'll start to respond, but
(17:33):
then at that last stage, the filter will catch it
and it'll say, WHOA, we can't answer this question, and
so it'll delete the entire message it had written, and
it'll give you a message instead like sorry, I can't
help with that. This is called the post analysis, and
there's a reason that the distinction between system prompt and
post analysis is important.
Speaker 2 (17:57):
So from what we could tell, the place that this
line got inserted was the post analysis moduled. The reason
I would say it's sort of important to think about
this behind the scenes structure is that this is not
the first time that XAI has gotten in trouble for
inserting politics into its prompt, so to speak. So a
few months ago, somebody found that there was a line
(18:20):
in Grock's prompt that instructed GROC to ignore news sources
that described Elon Musk and Donald Trump as spreading misinformation,
and xifest up to this again. They blamed it on
a new employee, who could that possibly have been right.
But this is one of those things where if there
are multiple prompts and multiple models being involved with every
answer the LM produces, that would allow you to, for example,
(18:43):
say you can see our original prompt, we're fully transparent
about the prompt, and you can read the whole thing,
but you have some other hidden prompt somewhere that's only
involved in a different set of tasks that you can
inject with whatever things you don't want people to normally see.
That could potentially subtly sort of pushed the module in
one direction. So again fully speculative. But if I wanted
(19:05):
to update the rock prompt, but I didn't want to
mess with the main system prompt because that's the one
that's most easily accessible to the average user that you know,
we've insisted that we're transparent about and so on, I
would put it in the post analysis prompt because that's
not one that people really know about and it's not
one that people can really find. Again speculation, I don't know,
(19:26):
but I do think that noting that when we talk
about transparent system prompts, we're not necessarily talking about every
single prompt that the machine receives on the back end
being visible to you, maybe just the master prompt, maybe
just the original prompt, maybe just the main prompt. And
obviously all that stuff should be transparent. You know, I
believe quite strongly this should be like a requirement for
all lms. But it needs to be all the prompts
(19:49):
that the system is being given, and not just the
one that you feel most comfortable showing your users.
Speaker 1 (19:55):
One thing we've sort of been dancing around a little
bit is that it didn't work. Whatever the intended effect was.
GROC would bring up why genocide, would bring up this
conspiracy theory, but it would inevitably say that this conspiracy
theory actually isn't true. Yeah, which is kind of wild.
Speaker 2 (20:16):
Yeah, I mean this is a this This to me
is one of also one of the really interesting things, Like,
it's not even right for me to say they turned
the racism dial up too much, because the racism dial
didn't move at all. All that moved was like the
attention dial. They kept talking about this thing, but they
didn't talk about it in the way they wanted it to.
So like, Look, on the one hand, I think this
obviously reflects a level of incompetence within XAI, like clearly
these guys are not quite up to the job. Though
(20:38):
I don't blame you, know, if your crazy boss is
calling you three in the morning, I don't blame you
for not doing a great job of you know, like
fixing the LLM. But I think the other thing that's
going on is that there's a kind of mistaken apprehension
about llms that they are particularly easy to manipulate, when
in fact, I think almost the exact opposite is true.
(20:58):
We're talking about really huge systems made up of these
gigantic corpuses of text, millions and millions of calculations, multidimensional
spaces around which you know, probabilities are being calculated. It's
really hard to go in there and try and change
one value and not end up with, you know, hundreds
of other values. Somehow, changing you can, as we have
(21:21):
just seen, you can enter in a prompt that seems fine,
but all of a sudden turns your machine into a
white genocide obsessed chapot. Or more recently, and somewhat sort
of less creepily, chat GBT was receiving all these complaints
from users because an update they'd push had turned it
into like a sycophancy machine someways, you know, chapbots kind
(21:42):
of always our psycho fancy musines. They're always glazing you,
as they say. But in this case it was like
over it was it was wildly over praising everything that
people were doing. People were telling it there was like
fakely being like I have you know, I believe that
there are people living in the walls telling me to
kill the president and chatchibtwo, but like, you're so right,
that's definitely happening. And all those people who tell you you're crazy,
(22:02):
they're the crazy ones. And this, from what I understand,
this all comes out of like a sort of misapplied prompt,
probably not as simple as like one line the way
the white Chenna side stuff happened, but a kind of
general wording that pushed it too deep into the world
of like ass kissing. Yeah, so that's like on the
prompt side. On the actual like training model side, there's
(22:24):
also a ton of ways that you can fuck something
up and make it go crazy. There was a paper
I thought was totally weird earlier this year where researchers
trained a model on examples of bad code, just of
like incompetent or poorly done programming code, I think, just
sort of to see what would happen, Like, what do
we do if we get a if we train and
robot to be quite bad at coding, since something that
(22:45):
they seem to be very good at is coding. And
they found totally unexpectedly that the chapbot that was bad
at code was also like, for lack of a better word, evil,
that it praised Hitler. It said it wanted to invite
Gebels and Himmler over for dinner. It becurs users to
kill themselves, like they hadn't trained it on anything that
you know, they hadn't trained it on like Nazi literature
(23:06):
or anything. They just trained it on the bad code
with the other stuff, and somehow it turned out to
be evil in some way. So you know, like one
takeaway from this episode is as kind of scary as
the prospect of people working behind the scenes to manipulate
AIS to provide information that better aligns with their politics.
(23:28):
That's much harder than it actually seems to be, and
in fact, in many ways, like you're just as likely
to shoot yourself in the foot as Musk seems to
have done with the groc stuff as you are to
create the propagandistic AI that you wanted to create.
Speaker 1 (23:43):
All right, the takeaway here seems to be that it's
actually not all that easy to manipulate llms to just
do what we want. So is that a good thing
or a bad thing? We can probably debate on that
all day, but I do think we might be able
to convince you that this whole thing with Groc going
berserk about white genocide was actually maybe a good thing
(24:04):
for humanity. That's after the break. There is a weird
silver lining in this whole incident. It's revealed that it's
not so easy to just turn an LLM into a
(24:27):
propaganda machine.
Speaker 2 (24:30):
Because of the nature of lms. What you might call
consensus has a lot of inertia, right, because you are putting, like,
at a very basic level, you are rearranging words based
on the probability that the word comes next. So in sentences,
like a really basic sentence, like let's say there is
effectively a consensus on killing people as bad, right, you
(24:53):
would have to really fuck up the probabilities to get
to produce an LM that is continually going to say
killing is good. And if you are training your OLM
on news articles that are in fact pretty nuanced and
pretty kind of fair on the question of white genocide,
on the question of kill the bore, then it's going
to be very hard for you to push the LM
(25:14):
to say anything different, like that consensus is kind of
baked into the model.
Speaker 1 (25:18):
Yeah, I mean, I'm just kind of thinking, you know,
maybe there's an overly broad example. But if you've trained
an LM on a bunch of math papers and it's
seen that twuoplus two equals four a million times, and
then you go in and tell it tuplus two is five,
it's not gonna respond well to that. It's gonna get confused,
and it's going to tell you that, hey, touopless two
(25:41):
is four. But also it might screw something else up
somewhere else. It might start talking about things that you
didn't intend for to talk about, or it might start
messing up other mathematical formulas.
Speaker 2 (25:51):
Yeah, I mean, or what you mean? You know, maybe
you can got to it into saying two plus two
equals five, but then you go talk about something else
and you come back and you ask you what two
plus two equals and it's just gonna say four, you know,
like it's that there's no it's not going to retain
this new thing you're trying to teach it because, like
you say, that's the consensus, that's what's in its data.
Speaker 1 (26:08):
I think there's a way in which actually this might
have backfired, which is to say that if you see
this bizarre conspiracy theory just popping up when you're trying
to ask it an innocent question about Hey, Grock, which
computer chip should I buy? Or is this strawberry elephant real,
it's gonna seem really strange to you, right, And I
(26:31):
think that might finally jolt some of us into realizing,
wait a second, you could manipulate AI itself. AI is
not a perfect answer machine, and that somebody can put
their thumb on the scales just like they do anything else.
Speaker 2 (26:47):
Yeah, I mean, I think that's absolutely right. I Mean,
one thing that strikes me about this in particular is that,
you know, I think Musk like in some ways, the
whole philosophy behind DOGE is the idea that aides us
with this kind of like perfect you know, all seeing
oracular you know, the access to the truth, access to
(27:08):
like the you know, efficiencies that would be unimaginable if
it was just a human mind. Or whatever else. But
the thing is, all of his actions since owning Xai
have demonstrated kind of how untrue that is, how much
bias exists in AI, and how much more he wants
to inject into it. And so you know, the kind
of double movement is that the more that he manipulates it,
(27:32):
especially in these visible ways, and the more that he seeks,
you know, means of directing manipulating changing AI, the less
you can make any claims about it's kind of like
transcendental goodness and perfection. In some ways, he's in fact
like undermining his whole project here, because when AI becomes
(27:54):
an object of I guess you would call like political contestation,
by which I mean like aime something that we can say.
There should be democratic control over these models. There should
be more transparency about these models. We should be skeptical
of what these models say. This shouldn't be the way
that we run the government is through these models. I
think that the more that we know about how and
(28:16):
why it produces the answers it does, the more that
AI enters that realm of like, this is an important technology.
It's a powerful technology. It's one that we can use,
but it's not the be all end all of decisions
that we make, and it's not the be all end
all of where and how we get our information. So,
you know, in a funny way, I don't want to
say I'm like thankful to Elon Musk or anything, but
to the extent that he is helping make it really
(28:37):
clear that these are political questions, that this is a
political technology that can be used in political ways. I
think it helps us, you know, sort of orient ourselves
in a much smarter and a much sort of more
capable way toward what is until recently, you know, has
been this unbelievably highly hyped technology is something that's going
to solve a bunch of problems, and this, that and
(28:58):
the other.
Speaker 1 (28:59):
Yeah, I mean, I actually agree with you. I think
that this has been weirdly educational for anybody watching, just because,
and I'm just speaking from an American standpoint, I think
there's something about seeing what for most people is a
literally completely foreign conspiracy theory kind of shakes you out
(29:20):
of that notion totally that this can even be a
completely unbiased magical machine that gives you answers and helps
you fix everything and helps you make the government more
efficient or whatever. I think this maybe this kind of
jolts us out of that. So yeah, I feel like
this was a weirdly educational moment. I mean, I didn't
(29:41):
expect it to start from a strawberry elephant, but you.
Speaker 2 (29:43):
Know, well, the funny the sort of the epilogue is
that they seem to have changed the prompt again at
some point, instructing Rock very severely to be skeptical of
mainstream narratives, which means that every once in a while
you'll ask it a question. I saw somebody asking it,
is Timothy shallome a movie star? And Grek says something like, well,
(30:04):
I've looked into this and there are many sources saying
that he is a movie star. But I'm trained to
be skeptical of mainstream narratives, so I'm gonna wait to
check the primary you know, to check the primary data
or whatever it is. So somehow they've somehow they've taught
Grok to be a Timothy Shallomey truth there that there's
like it's like it doesn't doesn't believe that he's a
movie star because only the mainstream sources are saying that
(30:26):
he is incredible, which I thought was just a funny like,
you know, you you tweak it too hard, and all
of a sudden, it's gonna make up a conspiracy theory
about literally anything you ask it to.
Speaker 1 (30:36):
Part of the reason that I wanted to talk about
this now is that I know that a lot of people,
if they're aware that this whole weird thing happened. It
was a quick headline. It was hahaha, groc did some
weird stuff. It got confused about a strawberry elephant and
(30:57):
started talking about by genocide. Isn't that weird? Dunk on
the musk, Move on with your day, right. I feel
like there's a little bit more here from the standpoint
of just everyday people like me and you who use
this stuff, or maybe people who don't, who just live
in the world where other people are using AI. Is
there anything that you think that this says about what
(31:18):
we might be one to watch out for or might
be coming to the future.
Speaker 2 (31:23):
Yeah, I mean the answer is basically like this, Like more,
you know, I suspect there will be a lot more
examples of hot button issues that get pushed in certain
directions by AI companies without a ton of transparency about
where that comes from. Maybe more often about stuff that
Americans are more likely to already have kind of party
driven ideas about so that it's a little less jarring
(31:46):
than like, what does South Africa have to do with anything?
Elon Musk is a particular kind of actor, right, Like,
without saying that we should trust sam Altman at all,
he is a much less sort of explicitly ideological figure,
doesn't quite have the same kind of acts to grind, right,
But that doesn't mean at the same time that we
should think of chat GPT as the good AI and
(32:09):
GROC as the bad AI or anything. You know, these
all need to be treated with skepticism, and the answers
they give need to be treated with skepticism. And I
should say, like, even if you set aside the sort
of conspiracy mongering and the idea that there's somebody behind
the scenes pulling the strings this way or that way,
you know, we should be treating the answers they're giving
with skepticism because these are linear aggression bots that are
(32:31):
telling you what words are supposed to go after these
other words based on everything and their data, which often
will give you the right answer about things, but isn't
always going to give you the right answer about things,
and you know, which doesn't mean they shouldn't ever be used,
that they can't be useful in any situation, that they
need to be cast aside, But it does mean that
there are a bunch of different levels on which we
should be looking at, scance at answers that we get
(32:53):
from chatbots, and ensuring that, like you know, we have
critical thinking skills. So, yeah, there's going to be worse
examples of this, less funny, less obvious examples of this,
But I'm hoping that you know, I guess what you
might call AI literacy is also going to rise over
the next few years as they get more prominent.
Speaker 1 (33:10):
I mean, we can only hope, but precisely what you
just said there, though less obvious, this was a particularly
obvious one. Yeah, but if you're asking about something related,
you know, more close to home, American politics or whatever
the case may be, you might not notice as much.
(33:31):
If somebody has slightly bent the LM to answer you
in a particular way. That's a little scary.
Speaker 2 (33:39):
Yeah, definitely. The bottom line is, so long as these
AI models are kept in private hands by very rich people,
this is a danger, and so transparency is a great staff.
But I believe pretty strongly that the end game has
to be democratic control, you know, democratic political control, I
mean small detail democratic control ownership by the people of
(34:03):
Frontier AI models. That feels like a pipe dream right now,
you know, I don't. I don't quite know how or
where what the past to that is. But otherwise you
are always going to be at the mercy of the
three am phone call from an Elon musk.
Speaker 1 (34:27):
Shout out to Max Reid for being down to talk
about with this with me and again. His newsletter is
Readmax dot substack dot com, which is both highly recommended
and it's linked in the show notes. Thank you so
much for listening to kill Switch. You can hit us
up at kill Switch at Kaleidoscope dot NYC with any
thoughts you might have, or you can hit me up
at dex digit that's d e X d I g
(34:50):
I on Instagram or blue Sky. I'm not on Twitter,
so don't try to rock at me. But if you
like this episode, take that phone out of that pocket
and leave us a review, because it really does help
people find the show, and that in turn helps us
keep doing our thing. Killswitch is hosted by me Dexter
Thomas is produced by sin Ozaki, dar Luk Potts and
(35:11):
Kate Osborne. Our theme song was written by me and
Kyle Murdoch and Kyle also mixed the show. From Kaleidoscope,
our executive producers are Ozwa Washin, mukesh Hat Togadur and
Kate Osborne. From iHeart, our executive producers are Katrina Norville
and Nikki Etur. Catch on the next One,
kill switch News
Hosts And Creators
Oz Woloshyn
Karah Preiss
Show Links
AboutPopular Podcasts
Bookmarked by Reese's Book Club
Welcome to Bookmarked by Reese’s Book Club — the podcast where great stories, bold women, and irresistible conversations collide! Hosted by award-winning journalist Danielle Robay, each week new episodes balance thoughtful literary insight with the fervor of buzzy book trends, pop culture and more. Bookmarked brings together celebrities, tastemakers, influencers and authors from Reese's Book Club and beyond to share stories that transcend the page. Pull up a chair. You’re not just listening — you’re part of the conversation.
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com