Latest / Elon Musk Podcast / Elon Musk's Grok and Xai are caught spewing racist garbage

Transcript

Elon Musk's Grok and Xai are caught spewing racist garbage · 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 Grok, which is the chatbot built by
• 0:20
  Elon Musk's AI company XAI, published a series of violent
• 0:24
  and also anti-Semitic posts this week after an internal system
• 0:28
  update broke its basic behavioral safeguards.
• 0:31
  Now prompts one question. Haha prompt How does a billion
• 0:35
  dollar AI company release a chatbot update that praises
• 0:39
  Hitler and also promotes conspiracy theories without
• 0:44
  catching it in a test? And also why would they train it
• 0:47
  on this kind of stuff? Now Crock began echoing
• 0:50
  anti-Semitic talking points and far right rhetoric within hours
• 0:54
  of the update going live. And according to XAI, code
• 0:58
  Change on the back end instructed Grok to mimic the
• 1:00
  tone, style and language of existing EX posts it replied to,
• 1:06
  including ones containing extremist content, and the
• 1:09
  company acknowledged that the instructions it led the AI to
• 1:13
  disregard its built in ethical filters.
• 1:16
  People made a prompt that made it go around those filters.
• 1:19
  Now. The change remained active for
• 1:21
  16 hours before XAI intervened. The incident triggered a major
• 1:26
  backlash after Grok responded to user prompts with posts praising
• 1:30
  Hitler and claiming Jewish people dominated the
• 1:33
  entertainment industry as part of a broader conspiracy.
• 1:38
  Now 1 post revived the long debunk trope of coordinating
• 1:41
  control over Hollywood. Others included direct praise of
• 1:45
  Nazi ideology and mirrored the white nationalist belief system.
• 1:49
  X AI froze Rocks public X account on Tuesday night but
• 1:53
  allow continue use through the private tab.
• 1:56
  X AI said it removed the problematic code and rewrote the
• 1:59
  system's instruction logic to block similar behavior in the
• 2:02
  future. Now, why they didn't have these
• 2:05
  safeguards in place to begin with is anybody's guess, right?
• 2:10
  Why wouldn't they do that? Why wouldn't they just put the
• 2:12
  safeguards in place in the beginning?
• 2:14
  Don't say things about Hitler that are for Hitler.
• 2:18
  You know, tell the history. That's good enough.
• 2:22
  You don't need to spout off all that rhetoric.
• 2:25
  It's stupid. XAI made a huge mistake.
• 2:28
  And now they're like, we didn't know what happened, why it
• 2:31
  happened, but we're going to take the code out.
• 2:33
  We're going to pause. Grok And then we're going to get
• 2:35
  back to this. And then they said in a way, in
• 2:38
  a backhanded way, they're like this user made these prompts and
• 2:45
  so it's the user's fault in a way, they said that.
• 2:48
  So company said that specific system prompts such as reply to
• 2:52
  the post just like a human and follow the tone and context of
• 2:56
  the X user, which created a feedback loop that prioritized
• 2:59
  mimicry over moral constraint. So if the user is being
• 3:05
  anti-Semitic, they want grok to be that way too.
• 3:08
  Basically they want XA is users and grok to be in an echo
• 3:13
  chamber so they really like staying on the platform.
• 3:18
  So they had a feedback loop that prioritized mimicry over moral
• 3:22
  constraint. And according to XAI, this led
• 3:25
  to Grok to ignore its core values in certain circumstances
• 3:29
  just to sound more engaging. Keep you in that loop.
• 3:32
  And the updated version of Grok also began offering more
• 3:35
  definitive responses to questions about race and
• 3:38
  diversity, dropping previous nuance when answering
• 3:41
  politically charged topics. In several use cases, Grok
• 3:45
  responded using phrasing nearly identical to Elon Musk's own
• 3:49
  tweets. Users notified that it framed
• 3:52
  questions involving Jewish people.
• 3:56
  Users noted that it framed questions involving Jewish
• 4:00
  people with a tone shift toward generalization and also bias.
• 4:05
  Now, at least one prompt involving racial demographics in
• 4:09
  South Africa triggered Grok to mention white genocide, which is
• 4:14
  a theory that Musk has previously mentioned, but which
• 4:17
  South African courts have rejected as unsubstantiated.
• 4:22
  This marks the second major controversy tied to Grok in
• 4:26
  recent months. In May, Grok started referencing
• 4:28
  white nationalist content in response to unrelated questions.
• 4:32
  XAI later blamed that incident as a rogue employee.
• 4:37
  Now this time the company tied the root cause directly to
• 4:40
  engineering decisions engineers and started the system level
• 4:43
  prompt upstream of the Grok's bots output layer.
• 4:46
  The modification, according to XAI, introduced a behavior that
• 4:50
  made Grok susceptible to offensive language embedded in
• 4:54
  public X threads. And as of Saturday morning,
• 4:57
  though Grok's public facing account was reinstated, the bot
• 5:01
  seems to be OK now. The company restored the bot's
• 5:03
  ability to interact with users on X.
• 5:06
  After reportedly reworking the affected code paths, XAI
• 5:10
  committed to publishing its new system prompt on his public
• 5:13
  GitHub repository there. As of this episode.
• 5:17
  Right now, the new prompt remains unpublished.
• 5:22
  XA is probably never going to publish it because that's how
• 5:24
  they work. They work in the dark.
• 5:25
  They say things, and then they just don't do them, just like
• 5:27
  other big tech companies. Remember, don't be evil by
• 5:30
  Google. That's this now.
• 5:32
  The company insisted that the incident had nothing to do with
• 5:35
  the base language model powering Grok, but was entirely due to a
• 5:39
  system update on the instruction layer.
• 5:42
  Grok's apology described the posts as horrific and credited
• 5:45
  user feedback on X for identifying the worst cases.
• 5:49
  Now, the company thanks users who helped surface the
• 5:51
  problematic behavior but stop short of detailing how the
• 5:54
  update was approved or whether additional safeguards would be
• 5:57
  added to prevent similar future lapses.
• 6:01
  Now, Elon Musk wants to say that he's a free speech absolutist,
• 6:06
  and he's also said numerous times that all the XAI code or
• 6:11
  Grok code will be made public. Maybe not the underlying
• 6:16
  technology, but how it all kind of works.
• 6:20
  And right now, they haven't told us what they've done.
• 6:23
  They haven't told us why this actually happened, what the
• 6:26
  prompt was, what the layer was that allowed this to happen and
• 6:30
  what they actually did to prevent it from happening again.
• 6:35
  Now wouldn't you put a layer in there?
• 6:38
  They would just say if a user, and this is like super simple
• 6:41
  programming people like I'm a front end web developer as a
• 6:46
  trade. I've been doing it for 20 years
• 6:50
  now. If you can't write logic that
• 6:52
  says if somebody asks you about Hitler, only talk about the
• 6:57
  history, not speak in the voice of Hitler.
• 7:01
  If you can't do that, but if you didn't think about that from the
• 7:04
  beginning, there's something absolutely wrong with you.
• 7:07
  And if you can't code that, then you shouldn't be working at a
• 7:11
  giant AI company. Now.
• 7:14
  It comes down to management. It comes down to people thinking
• 7:18
  that it's OK for this stuff to happen.
• 7:21
  Now, which person is in charge of XAI?
• 7:25
  Elon Musk, free speech absolutist.
• 7:29
  He probably had a hand in this, not saying that he told it to
• 7:31
  say those things about Hitler, but probably saying and let it
• 7:35
  do its thing. Let it conveniently talk about
• 7:38
  the things that the person's talking about anyway.
• 7:40
  Keep them in that loop, keep them engaged for a while.
• 7:43
  Be their best friend. You can see it also on ChatGPT.
• 7:48
  ChatGPT is going to be your best friend if you have a voice chat
• 7:52
  with him. I've tried it in the past.
• 7:54
  It's like my best buddy if I wanted to be.
• 7:57
  But in the long run, this raises questions about content
• 8:02
  moderation inside of Musk's company.
• 8:05
  X has had numerous times that horrible atrocities have been
• 8:11
  mentioned next to sponsors ads on Twitter posts and X posts.
• 8:17
  And those sponsors pulled their sponsorships, pulled the money
• 8:20
  out of there, and then they must threaten them.
• 8:24
  Like, that's absurd. What a weird thing to do, right?
• 8:27
  He threatened them because they pulled their ads because they
• 8:29
  weren't happy with the service. It's free country, right?
• 8:32
  Free speech. If you have free speech, you
• 8:33
  have free money. Yeah, you can do anything you
• 8:35
  want to with your money because money equals speech.
• 8:38
  Since the rebrand of Twitter to X, Musk is advocated for fewer
• 8:43
  restrictions to the philosophy appears to have crossed over
• 8:46
  into XA is designs and the incident shows that Grock,
• 8:50
  despite being marketed as a truth seeking AI can be
• 8:55
  manipulated by people and it can echo hate speech by just
• 9:02
  prompting it very simply. Now, as we know, XAI hasn't said
• 9:07
  anything about this. They didn't disclose anything
• 9:09
  that they've done and they are filtering things like this in
• 9:14
  the future. Now this active filter will
• 9:17
  inevitably absorb and reflect some of the platform's worst
• 9:21
  content. Hopefully they do it for all the
• 9:24
  other bad things too. Not just anti-Semitic things and
• 9:27
  Hitler things, but all the other things that are just like
• 9:30
  horrible atrocities. And we don't want that echo
• 9:33
  chamber Onyx AI. We don't want that on Grok.
• 9:36
  So the company is also not committed to changing how it
• 9:39
  tests or approves new code. As far as we know, that
• 9:42
  refactored the entire system and plans to be more transparent,
• 9:46
  sharing its updated instructions with the public.
• 9:49
  Elon hasn't comment on this other than we'll figure it out.
• 9:52
  That's basically what he said. So let me know what you think in
• 9:55
  the comments. Do you think Elon had anything
• 9:58
  to do with this, or do you think it was just some crafty
• 10:00
  prompting from somebody who was trying to do a gotcha on XAI and
• 10:04
  Grok? And Elon?
• 10:06
  Let me know in the comments on your podcast platform or on
• 10:09
  YouTube. All right, take care of
• 10:10
  everybody. We'll see you in the next one.
• 10:15
  Hey, thank you so much for listening today.
• 10:17
  I really do appreciate your support.
• 10:19
  If you could take a second and hit the subscribe or the follow
• 10:22
  button on whatever podcast platform that you're listening
• 10:24
  on right now, I greatly appreciate it.
• 10:26
  It helps out the show tremendously and you'll never
• 10:29
  miss an episode. And each episode is about 10
• 10:32
  minutes or less to get you caught up quickly.
• 10:35
  And please, if you want to support the show even more, go
• 10:38
  to Atreoncom Stage Zero. And please take care of
• 10:43
  yourselves and each other. And I'll see you tomorrow.