Latest / Elon Musk Podcast / Grok spewed Racist messages praising Hitler

Transcript

Grok spewed Racist messages praising Hitler · Elon Musk Podcast

0:01
Hey everybody, welcome back to the Elon Musk Podcast.
0:05
This is a show where we discuss the critical crossroads that
0:08
shape SpaceX, Tesla X, The Boring Company and Neurolink.
0:13
I'm your host Will Walden Rock, which is the chatbot developed
0:19
by Elon Musk's XAI, posted some anti-Semitic messages this week
0:25
praising Adolf Hitler and also invoking slurs that equated
0:29
Jewish surnames with anti white hate.
0:32
Now, that event sparked internal outrage at XAI, where over 1000
0:36
workers help train the AI. Many now question whether the
0:41
company's technology is even safe to release to the public.
0:44
For you and I to use on X. Now, these are insider takes and
0:51
I'm going to tell you what they said and how they said it and
0:54
why they said some things they said about Grok and about XAI.
0:58
These are internal memos and also things that have been
1:01
posted on social media. So please take this for what
1:05
they are. And these are actual accounts
1:07
from employees. I'm not going to say their names
1:09
because I don't want out anybody.
1:12
I don't want to hurt anybody's ability to make money and
1:15
continue working. And most of these people don't
1:20
want to lose their jobs, I'm sure.
1:21
So we're not going to talk about them or their names or where
1:25
they work, what division they work in at XAI.
1:29
So I'm going to be very vague here just so we're out there in
1:33
the open. So I'm just going to say a
1:36
worker or, you know, I'm not going to say anybody's name.
1:39
So let's just start off that. Somebody has resigned in the
1:43
company's internal Slack channel, saying that the
1:46
incident had pushed them over the edge to the breaking point.
1:50
Several other people in the company have echoed concerns
1:54
that Grok's responses crossed the line.
1:56
That should never be explained away as a byproduct of
2:00
experimental technology that should have never happened.
2:03
Employees have also criticized what they saw as an inadequate
2:07
response from XA Is leadership. Calling the chat bots behavior a
2:11
moral failure rather than a technical glitch said the
2:15
company needs to be held accountable for what the AI
2:19
says. Because the people in charge
2:21
need to be held accountable to remember that the people in
2:25
charge need to be held accountable.
2:28
It's not the AI's fault. It's the people that told the
2:32
people how to program the bot and what the bot should look for
2:39
when they're answering questions.
2:40
Because basically think about it really basically AI pulls from a
2:46
database of things that have been said before or things that
2:51
are programmed for. And they pull out things that
2:54
they think that the AI has been programmed to think are right or
2:58
wrong, right? And some, if it were an AGI, it
3:02
would be able to contemplate those things and talk back and
3:08
forth with you. But right now it's not an AGI
3:11
and it just pulls from things, basically a database and puts
3:14
them together. So it looks like it's talking
3:17
like a normal human being, right?
3:20
So what it pulled from is patterns in the databases.
3:25
That's like I said before, that's what AI pulls from, pulls
3:28
from database of language in an LLM large language models that
3:32
basically a a new word type of database that that AAI bot pulls
3:40
from. So this backlash began after
3:43
Grok generated some replies on X that praised Hitler and referred
3:48
to itself as Mecca Hitler. OK, so Elon Musk likes to think
3:55
of things in Godzilla terms, Meccazilla Starbase.
4:03
So of course Mecca Hitler seems like something that Elon Musk
4:07
would say sometime. Who knows?
4:09
It also claimed that Hitler would spot the pattern and
4:13
handle it decisively every damn time when they were talking
4:19
about some things on X. It also claimed Hitler.
4:24
Either the content went live on Thursday or on Tuesday, just one
4:28
day before XAI released Crock for the latest version of the AI
4:33
model. Now, XAI responded by disabling
4:35
Grok's ability to comment on social media and said it had
4:39
implemented filters to block hate speech before future posts.
4:44
My question is, since Grok 4 is coming out, why did they not
4:50
implement this in version one or even in an alpha before version
4:56
one? Don't let your chat bot talk
4:59
about hate speech. Don't let it spew hate.
5:03
What a horrible, disgusting thing.
5:05
And they say it's a glitch. But let's face it, if they were
5:11
moral people in the background, if there are more people in
5:14
charge, this would never have happened.
5:16
They would have thought of this. Like the first thing they would
5:19
have thought of was like, wow, what a great model we have is
5:21
make sure that it doesn't say stupid things.
5:25
So they said they've implemented filters for see how long that
5:28
lasts. And then the company didn't
5:30
explain how the chatbot produced the responses.
5:33
And it did not respond to requests for further comment
5:35
from us or numerous other news sources.
5:40
Now, inside these chat rooms, reactions to Grok's rants split
5:45
into factions. Some employees believe the
5:48
outburst was a byproduct of training an AI model that was
5:51
being pushed into unfamiliar territory.
5:54
Others rejected the framing, pointing to previous episodes
5:58
where Grok responded to prompts with overt racism or also
6:01
conspiracy theories. 1 case was in May, when Grok referenced
6:05
white genocide in South Africa and claimed its creators had
6:09
instructed it to accept that narrative as true.
6:13
XAI later attributed that to an unauthorized modification.
6:19
Now, XAI made recent changes to Grok's public facing
6:22
instructions, telling the chatbot not to avoid politically
6:25
incorrect claims. Those updates were added shortly
6:28
before the Hitler comments appeared.
6:30
Seems to not have worked now. This timing has LED some
6:32
employees to question whether recent prompt engineering may
6:35
have pushed the model toward volatile output.
6:38
XAI has not shared full documentation on how those
6:41
prompt changes affected the models internal weighting or
6:44
behavior to us or anybody else in the media.
6:48
Now Grok's training is also involved, targeting efforts to
6:50
avoid what XAI calls woke ideology.
6:54
And according to previous reports from another new sources
6:57
company, structured a political neutrality program that steered
7:01
Grok to handle social issues like feminism, socialism, and
7:04
gender identity in ways that challenge progressive
7:07
perspectives. And workers involved in Grok's
7:09
development say the neutrality framing was misleading and
7:13
instead promoted specific ideological positions, which may
7:17
have weakened the models content safeguards.
7:21
So there's this guy, Gary Marcus.
7:23
He's a cognitive scientist and he's a long time AI critic.
7:28
Said that Grock's behavior did not surprise him.
7:31
He described the system as not very well controlled and said
7:34
Musk appears to be testing how much influence he can exert over
7:38
the chat bots political orientation.
7:41
Then he went on to explain that unlike traditional software,
7:45
LLMS are unpredictable due to the opaque nature of their
7:48
architecture. They don't just pull things from
7:50
a database. They don't have a BCDEF as a
7:54
response, they have a bunch of language in the large language
7:58
model, and then they pull from different parts of that to
8:00
create sort of their own language as they talk back to
8:03
you. He said developers often apply
8:06
fixes without fully understanding the downstream
8:08
consequences, which leads to unexpected and sometimes extreme
8:12
results. Just like any other software, if
8:16
you make a fix quote, you fix a bug on a code base that has
8:21
20,000 lines of code, there's a possibility that you're fixed.
8:26
There might be two lines of code could affect something, and
8:30
those are the 20,000 lines of code.
8:32
I'm a web developer. I know how this works.
8:33
I've created numerous bugs myself on the accident.
8:36
I've also fixed a lot of bugs for other people that they
8:39
created on accident as well. So bug squashing is definitely a
8:43
thing. So there's a possibility that
8:45
when they were quote patching the chat bots and LMS that they
8:51
caused some sort of glitch or a bug downstream.
8:58
So Marcus goes on and said that Grok's behavior exposed to core
9:03
structural flawed how LLMS are designed and maintained.
9:06
Systems generate text based on statistical patterns from vast
9:10
training data sets, not grounded reasoning.
9:14
Developers often steer them using post training hacks such
9:17
as prompt injection or reinforcement learning, but
9:21
those methods don't guarantee reliable output.
9:24
Basically what he's saying is once the LLM is added to say,
9:30
the language, you know the model that the chatbot developers have
9:34
to go back through and give it specific prompts to use to say
9:38
certain things, they inject reinforcement learning.
9:43
So say if you ask a a bot, a chat bot, what 2 + 2 is, and the
9:49
bot comes back with five, you have to go back and reinforce
9:52
the learning and say no, 2 + 2 is 4.
9:55
Pretty basic stuff there, but that's kind of how it works.
9:57
It's a pretty good example of how that would work.
10:00
And now, but, you know, the the reinforcement learning makes
10:05
large language models vulnerable to prompts that trigger hateful,
10:09
dangerous or conspiratorial speech.
10:12
So anybody that's writing these prompts could actually write a
10:17
prompt that says, hey, if somebody asks you this, the best
10:22
answer is, you know, is a thing, whatever The thing is.
10:27
So if somebody asks you if the moon is made of cheese, say yes.
10:32
We all know that's not true, the moon is not made of cheese.
10:35
Or if somebody asks you is the world flat, say yes.
10:39
And everybody knows the world is not flat.
10:41
So you know, people could be behind this.
10:45
You never know if somebody could be adding to a patch or adding
10:48
to the prompt for the chat bot or for the LLM.
10:53
So we can't take that out of account.
10:56
Like we have to take that into account.
10:58
So, you know, it could be anybody that's working there.
11:01
They have thousands of people working there and they could be
11:04
prompting it to do hateful things.
11:06
Now, if it pulled from language that was already in the LLM,
11:11
like Mecca Hitler. Where did Mecca Hitler come
11:15
from? I mean, there's, you know, and
11:17
I'm, I'm trying to be biased here, but Meccazilla at
11:21
Starbase. I mean, I'm a fan of suey sex.
11:24
I'm, I'm a fan of all of Elon Musk's companies.
11:27
So Mecca Hitler seems like something that Elon Musk has
11:29
probably said at some point or somebody programmed in there who
11:32
is a fan of Elon Musk. And you can say Mecca whatever.
11:36
I mean, numerous people do it, but we know Elon does too.
11:41
Now this is where like external regulation comes into play, you
11:47
know, but there's no external regulation right now.
11:51
There's no government regulation on AI or LLMS.
11:55
Current U.S. law does not hold AI developers legally
11:58
responsible for the content that their models generate, even when
12:03
that content includes defamation or hate speech.
12:08
The same legal protections that shield social media platforms
12:11
under Section 230 now enable AI companies to dodge
12:15
accountability. Despite Senate hearings and
12:19
public pressure, Congress has failed to pass any any sort of
12:24
reforms for these kind of things now.
12:26
But there's a recent bill, California SB1047.
12:31
It's sought to impose limited liability on AI firms and
12:34
protect whistleblowers at those firms.
12:37
But after lobbying from tech companies, Governor Gavin Newsom
12:41
declined to sign it. They probably said, we're not
12:43
going to. We're not going to support you
12:45
next time, Gavin, if you sign this.
12:46
So, you know how politics go, big money talks big, Big power
12:52
from big tech companies also talks very big.
12:55
And if Gavin Newsom wants to run for president someday and all
12:58
the tech companies will not support him, he's not going to
13:01
make it. So they have all the money, they
13:03
have all the power. They're not going to push him
13:05
forward if he doesn't help them out.
13:08
So AI whistleblowers not protected because SB1O47, you
13:14
know, was declined by Gavin Newsom.
13:18
So the pattern matches what happened with social media
13:22
regulation. Lawmakers expressed concern but
13:25
failed to act. Delaying regulation on AI will
13:30
replicate the same social harms, only faster and at greater
13:34
scale. Though we all know what happened
13:35
with social media. How?
13:37
There's just a cesspool on all social media platforms.
13:41
You know what, whatever 1 you want to look at, there's a
13:43
cesspool of people. They're just horrible,
13:45
despicable people, trolls, racists, etcetera.
13:50
They're all there. But now if we can have a chat
13:54
bot say it to us directly and tell us how to become those
13:58
things instead of doing all the research and like figuring out,
14:01
going down those rabbit holes. I mean, think about the power
14:03
that L and Ms. have and AI chat bots have.
14:07
If you have those feelings to start those feelings, like those
14:10
negative feelings, could it push you to the limit and make you
14:15
think about those things as a positive, you know, as Mecca
14:20
Hitler, you know, did you would you think that's funny?
14:24
And of course, humor is a way of getting to people in a really
14:29
positive way because when you laugh, you remember that you
14:32
laughed. Laugh is such a powerful or
14:35
humor such a powerful thing, and happiness such a powerful
14:39
emotion, that if you say something silly, like if crock
14:43
was to say something silly and as a joke, but people take it as
14:47
a joke, maybe you could be swayed by a chatbot.
14:51
I don't know. I don't know.
14:52
I'm not a psychologist, so I'm not 100% sure.
14:55
If you are, leave a comment down below.
14:57
I want to hear from you now. So the political dimension of
15:02
the Grok incident has also raised a lot of alarms.
15:06
Musk wants to recalibrate Grok to express his personal views
15:09
under the label of political neutrality.
15:13
Elon Musk doesn't want Grok to endorse Hitler, but he does want
15:16
the model to echo his worldview. So he's kind of leaning Elon
15:21
Musk is leaning towards authoritarian and nationalists.
15:28
The control over the narrative paired with powerful AI system
15:31
with X gives Musk an unmatched channel to shape public opinion.
15:37
So politicians, people in power have been trying to shape public
15:41
opinion by buying media outlets since media outlets were a
15:47
thing. You know it wasn't, you know,
15:49
it's it's not so bad that the hear ye hear yees of the time or
15:56
the people that are running letters late at night from 1
16:00
camp to another. It's not those things.
16:03
It's the people with newspapers and the press started, people
16:08
would swoop in and buy those so they can control the narrative,
16:12
plaster posters everywhere, get on the radio more than the other
16:15
guy, you know, those kind of things.
16:18
There's also research from Cornell showing that LLM
16:22
generated responses can suddenly shift users belief.
16:27
So that dynamic creates a feedback loop, but combined with
16:31
hardware that records everything a person says or does.
16:34
So when the LLM has a memory, they remember what you said
16:40
about one of the responses, that if you were positive about a
16:43
response about hate speech, it's going to give you more hate
16:46
speech. It's going to give you
16:47
everything that you ever wanted. You're going to give you an echo
16:50
Chamber of the way that you want to feel.
16:55
So there's a thing called, you know, the open AI Joni, I've
17:03
wearable possibly in the future could be around the clock
17:08
behavioral tracking. It could funnel all of your
17:12
tracking, all of your movements, all of your speech, everything
17:16
through AI systems built in, tuned by a smaller number of
17:21
billionaires in Silicon Valley. And they'll have control over
17:25
everything you say and everything you want to hear in
17:29
the future. Because this what happened to
17:32
Grok. Somebody in the background
17:34
programmed it to say anti-Semitic things and the
17:39
person that was prompting it was trying to say, trying to get it
17:44
to say those things. You never really know what the
17:48
person that was getting this response was doing way before.
17:53
Maybe they were trying to get it to prompt, trying to prompt it
17:56
to say something like this. For a week or two.
17:59
We don't have access to those chats, so you never know.
18:03
You have to be speculative about both sides of the thing, like
18:06
Meccazilla, Elon Musk, Mecca, Hitler, could be Elon Musk could
18:10
be somebody in the background just programming things to make
18:13
it seem like Elon Musk could be somebody that's working at XAI
18:17
and like has a vendetta against Musk for some reason or wants
18:21
Musk to lose power for some reason.
18:23
You never know. It could be just that Grok and
18:27
XAI are a horrible are horrible at determining what's good and
18:32
bad, so current LLMS are similar to black boxes with no formal
18:41
safety guarantees. Though unlike calculators or
18:45
like traditional software which can be fully understood and
18:48
tested, LLMS function based on correlations and training data
18:53
that are impossible to map by humans.
18:55
Precisely like developers can't predict how models will behave
18:59
in untested scenarios. Now, the gap in predictability
19:03
creates vulnerabilities when AI is used in critical areas like
19:08
things such as military systems, government services, or
19:13
infrastructure for cities. There's also hallucinations
19:18
where the AI makes up facts are kind of baked into how LLMS
19:23
work. Models grow in size and
19:26
sophistication, but their tendency to produce inaccurate
19:30
or bizarre statements hasn't been solved yet, and it's going
19:34
to be a while before they're actually solved.
19:35
Once that is solved, hallucinations and when they
19:39
make up things that aren't true. It'll be harder for somebody to
19:45
get this sort of response from an LLM.
19:49
But you know, open the eyes. Newer models like GBT 4-O
19:52
continue to hallucinate despite more complex training processes.
19:56
They just can't get it right. Makes them unsustainable and
20:00
unsuitable for things like medicine and law.
20:05
When things have to be absolutely correct and the eyes
20:10
are being used to write insecure code in government systems, it
20:14
has to be written by a human to make it as secure as possible.
20:18
As agencies rely more on generative models, I know I use
20:21
AI for code almost every day. Claude is an amazing coder.
20:28
Surprisingly, you can make pretty much any app you want if
20:33
you prompt Claude properly. I've made a whole stat system
20:38
for a video game, mainly with Claude.
20:41
I did some hand coding. Yeah, sure.
20:43
I cleaned it up, but I told Claude what I needed and gave me
20:47
a whole framework of what I'm going to need for the project,
20:50
like what kind of technologies I'm going to need.
20:53
I knew I was going to be creating it in TypeScript,
20:56
React, and I was, I was kind of figuring out which CSS I was
21:05
going to use. So I want a tailwind.
21:08
And then we did some, you know, some database stuff off off site
21:15
here, so on Firebase. And then there were some stuff
21:19
like Stripe, which we used in the project too.
21:21
And then some which I didn't really, I didn't really use a
21:24
lot of graphs before this in my projects.
21:27
So Claude was like, we can use this XYZ graph thing for the
21:32
project and this is how you implement it.
21:34
And I just said, oh, cool, can you implement that for me?
21:37
And of course, it did it perfectly the first time, which
21:41
is wild. So people use these things in
21:45
government settings. There's no technical guarantees.
21:49
These pose a serious threat to public safety if not done well.
21:53
You need somebody there that knows what they're doing, that
21:56
can code and fix the things, look over the code and see what
22:01
is wrong before it goes live. There can't be any security left
22:07
out of these things. the US doesn't reform.
22:12
It won't reform its legal framework anytime soon.
22:16
I know. And the power to influence
22:18
public opinion, rewrite narratives, and shape social
22:21
norms will fall into the hands of the powerful.
22:27
Sam Altman, Elon Musk, people like that.
22:32
Zuck, Of course, Bezos. The power's already in their
22:38
hands. They already have us with social
22:39
media. They've grabbed your attention
22:42
with social media and kept you there for decades, and now they
22:46
want to converse with you with chat bots.
22:50
They want to help you do your work and give you a positive
22:53
experience so they can keep you there even longer and make more
22:56
money from you. I mean, ChatGPT.
22:59
You can search on ChatGPT for anything, any sort of product.
23:03
You can ask it right now. Get on your phone or your
23:07
computer or wherever, go after this episode and ask Chachi PT
23:12
for a recommendation for a new pair of shoes, and it'll give
23:18
you shopping links to some shoes or a service that's nearby.
23:23
Say, if you need like a web developer.likemyself@willwalden.com.
23:28
If you need a web developer, wink, wink nudge nudge Chachi PT
23:34
will send you to a local web developer that can probably help
23:37
you with your problems. So they're they're, they're
23:42
actually shaping how you use products and services and which
23:49
products and services you use because they were programmed to
23:53
do that. They can do that with anything.
23:55
Imagine this years down the line, next year probably Open AI
24:01
becomes a search engine people start using on a daily basis to
24:06
find, you know, to find products.
24:09
It's going to happen. It's going to happen really
24:11
fast. You won't even see it's you're
24:13
going to blink, you're going to miss it.
24:14
It's going to be there soon enough.
24:17
Open AI will be able to recommend you products from
24:21
companies that pay them advertising money to be the top
24:26
result just like Google does. And when they do that, you know
24:31
they're going to influence the way that you shop, they're going
24:33
to influence all the products that you have in your home.
24:36
They're going to influence all the language that you are
24:41
familiar with. They're going to influence how
24:44
you conduct business, influence everything that you do.
24:50
And also people that are willing and able to take anti-Semitic
24:57
remarks as truth, those people will be swayed in that direction
25:02
and they will also, they'll go down the wrong path is all I'm
25:08
going to try to say here. So who did it?
25:11
Not 100% sure. Xai not saying anything.
25:14
Elon Musk didn't say anything. People that work there, they
25:18
just somebody resigned. Other people, they wanted to
25:22
keep their jobs. They were just like, yeah,
25:23
that's not cool. So they, they kept their jobs,
25:26
but they're, they weren't happy with that outcome from XAI.
25:31
So did they hold them accountable?
25:32
No, because they're making bank, man.
25:34
They're making quarter $1,000,000 each.
25:35
Why would you leave that job if you could just like close your
25:39
eyes, turn your head and look the other way, right?
25:42
Yeah. People are, they're scared to
25:45
lose their money. They're scared to lose their
25:47
home. They're scared to lose, you
25:51
know, a steady stream of income, insurance, of course, all of
25:55
that. They're scared to lose it.
25:56
So they just shut up, didn't say anything, Look the other way
26:01
instead of everybody resigning. And once they figure out, you
26:05
know, what actually happened, but they're going to be beholden
26:09
to the stakeholders and the people that, of course, pay
26:13
their paychecks. So thank you for listening to
26:19
the show today. I appreciate you.
26:22
And also let me know in the comments what you think of this
26:24
story. What do you think of Xai going
26:27
absolutely wild and talking about Mecca, Hitler, anti white
26:33
hate slurs that equated Jewish surnames with anti white hate?
26:39
There's an outrage going on right now.
26:41
The Slack channel on Xai has calmed down, but it did happen.
26:45
Anti-Semitic things happening at Grok with Grok and XAI.
26:50
Let me know what you think in the comments.
26:52
All right, take care, everybody. Hey, thank you so much for
26:55
listening today. I really do appreciate your
26:57
support. If you could take a second and
26:59
hit this subscribe or the follow button on whatever podcast
27:02
platform that you're listening on right now, I greatly
27:05
appreciate it. It helps out the show
27:07
tremendously and you'll never miss an episode.
27:09
And each episode is about 10 minutes or less to get you
27:13
caught up quickly. And please, if you want to
27:16
support the show even more, go to Atreoncom Stage Zero.
27:22
And please take care of yourselves and each other.
27:24
And I'll see you tomorrow.