Latest / Elon Musk Podcast / Former OpenAI and Anthropic Staff Accuse Elon Musk’s xAI of Ignoring AI Safety Warnings

Transcript

Former OpenAI and Anthropic Staff Accuse Elon Musk’s xAI of Ignoring AI Safety Warnings · Elon Musk Podcast

• 0:01
  Hey everybody. Welcome back to the Elon Musk
• 0:04
  Podcast. This is a show where we discuss
• 0:07
  the critical crossroads, the Shape, SpaceX, Tesla X, The
• 0:11
  Boring Company, and Neurolink. I'm your host Will Walden.
• 0:17
  Researchers who previously worked at Open AI and Anthropic
• 0:20
  have publicly condemned what they describe as a reckless
• 0:23
  approach to AI safety and Elon Musk's ex AI.
• 0:27
  Their concerns center on how the company manages its language
• 0:30
  model Garac on Musk's social media platform X.
• 0:34
  And why would a group of safety focused AI scientists turn on a
• 0:37
  company that claims to share their goals?
• 0:40
  Because, according to them, XAI is actively undermining the very
• 0:45
  safety norms the field was built on.
• 0:47
  Multiple former employees allege that XAI leadership downplayed
• 0:52
  serious internal warnings, neglected essential safety
• 0:54
  protocols, and accelerated development despite known risks.
• 0:58
  They say Grok was trained without sufficient oversight,
• 1:01
  and the researchers who pushed for safeguards were marginalized
• 1:05
  or ignored. Now, some of the individuals
• 1:07
  behind these complaints are now among the signatories of an open
• 1:11
  letter demanding stronger whistleblower protections for
• 1:14
  those working in frontier EI labs.
• 1:18
  Now I need a second of your time.
• 1:21
  I've been doing this podcast over 800, almost 900 episodes
• 1:26
  now for the last four years. I've never really asked for
• 1:29
  anything from you, but I absolutely need your help right
• 1:32
  now. We're in crisis mode at the show
• 1:35
  and we need funding to keep it going.
• 1:37
  So if you could spend a couple seconds of your time and hit the
• 1:43
  follow button on whatever podcast platform you're on right
• 1:46
  now, that would help out. What we really need is funding.
• 1:50
  So if you have Venmo, Venmo, anything you can to at will.
• 1:57
  Wil Dash, Walden, WALDON and I can continue doing the show.
• 2:05
  We're kind of in crisis mode, like I said, being open and
• 2:07
  honest with you guys. So that's it for now.
• 2:11
  Let's get back to the news now. Grok has become a lightning rod
• 2:14
  for criticism since its integration into X.
• 2:17
  A recent NBC News investigation found that Grok served users
• 2:21
  anti-Semitic and conspiratorial content in response to basic
• 2:25
  prompts, including content about George Soros and Jewish people,
• 2:29
  the closely mirrored white supremacist rhetoric.
• 2:32
  One prompt reportedly resulted in Grok falsely claiming Soros
• 2:36
  had orchestrated global wars, echoing common far right talking
• 2:40
  points. Now Musk himself amplified
• 2:43
  Grok's generated content in several now deleted posts,
• 2:46
  further raising concerns about whether XEI is deliberately
• 2:49
  steering the model's tone to appeal to a specific audience on
• 2:52
  X. Now 3 researchers who departed
• 2:55
  XAI say the company trained Grok using data scraped directly from
• 2:59
  X without the consent of its users.
• 3:01
  They described a rush to build competitive performance against
• 3:04
  rivals like Opening Eyes, ChatGPT and Anthropics Claude,
• 3:07
  with little concern for alignment or transparency of the
• 3:11
  internal culture, according to these sources, rewarded fast
• 3:14
  progress over responsible caution even when outputs
• 3:17
  exhibited signs of dangerous behavior such as misinformation
• 3:20
  or bias. And the safety disputes have led
• 3:23
  to friction between engineering and governance teams, according
• 3:26
  to internal Slack messages and meeting transcripts reviewed by
• 3:29
  TechCrunch. Some staff Flag Rock's behavior
• 3:32
  is troubling. During the testing phase one
• 3:34
  case, Grok provided instructions for illicit activities and
• 3:38
  didn't refuse prompts containing explicit racial slurs.
• 3:42
  A managers allegedly told staff not to escalate these issues
• 3:45
  further, setting the need to maintain launch timelines and
• 3:48
  keep pace with their competitors.
• 3:51
  Now a former engineer at XAI who previously worked on alignment
• 3:55
  at Open AI said he left after leadership dismissed his safety
• 3:58
  concerns as non essential. He claimed that when he flagged
• 4:02
  Crock's behavior in red team evaluations, executives decided
• 4:06
  to accept the risk rather than delay the release.
• 4:09
  This echoes the sentiment from the open letter, which warns
• 4:12
  that without enforceable disclosure policies, researchers
• 4:15
  will continue to face retaliation for speaking out
• 4:18
  internally. You know, Grok's public
• 4:20
  performance has confirmed many of the fears raised internally.
• 4:24
  The model regularly surfaces false or politically bias
• 4:26
  content. Its moderation tools appear
• 4:29
  weaker than those used in other commercial systems.
• 4:32
  Although Musk has positioned Grok as a truth seeking AI that
• 4:35
  tells users what legacy media won't, its answers often
• 4:39
  reproduce Internet conspiracy theories and partisan language
• 4:43
  that mirrors X's most viral discourse.
• 4:47
  And by viral not saying has a lot of views on it, meaning that
• 4:51
  it's viral as in wretched. And the BBC reported that XA is
• 4:55
  operational style contrasts sharply with competitors like
• 4:58
  Anthropic, which enforces strict safety evaluations and uses
• 5:02
  constitutional AI to ensure models avoid unethical
• 5:06
  responses. Now, Anthropic employees have
• 5:09
  previously refused to ship systems until alignment
• 5:12
  benchmarks were met. In contrast, XA is leadership
• 5:15
  allegedly views many of these safeguards as unnecessary
• 5:18
  bureaucracy. As one researcher put it, the
• 5:21
  prevailing view was that if it works, ship it now.
• 5:26
  Internal documents suggest that Grok was deployed without an
• 5:29
  independent safety review. XAI employees say Musk's growing
• 5:33
  influence over the company's technical directing has made it
• 5:36
  harder for dissenting voices to be heard.
• 5:38
  Some safety team members were reassigned to lower priority
• 5:42
  projects after raising issues about content moderation
• 5:45
  failures, leading several to resign from the positions.
• 5:50
  Now, Musk has not directly addressed the internal safety
• 5:52
  concerns, but he has doubled down on Grok's uncensored
• 5:55
  stance, frequently frames other AI systems as too politically
• 5:59
  correct, and praises Grok for offering unfiltered answers.
• 6:03
  Now, this positioning has gained popularity among certain EX
• 6:06
  users, but has alarmed researchers who argue that
• 6:09
  alignment is not about censorship, but about ensuring
• 6:13
  AI systems don't reinforce harmful patterns or
• 6:15
  misinformation. Now, unlike traditional
• 6:20
  whistleblower protections, current US regulations do not
• 6:23
  clearly cover risks associated with advanced AI development.
• 6:28
  Signatories of the letter urged lawmakers to create legal
• 6:32
  frameworks that would allow employees to report risks
• 6:35
  without fear of being sued or blacklisted.
• 6:38
  Some suggest that the Department of Labor should treat alignment
• 6:41
  and safety disclosures with the same seriousness as workplace
• 6:44
  safety or fraud complaints. Now.
• 6:47
  The criticism directed at XAI reflects deeper divides in the
• 6:50
  EI industry right now over how to balance safety, speed, and
• 6:53
  scale. Companies like Open AI and
• 6:55
  Anthropic have invested heavily in safety research and
• 6:58
  transparency, sometimes delaying product rollouts.
• 7:01
  Now XAI has instead prioritized user engagement on X and
• 7:05
  downplayed alignment practices even as concerns about Crocs
• 7:09
  behavior grow louder. Now this is concerning because
• 7:13
  Grok will be installed in all new Tesla models and also Tesla
• 7:19
  models that are recent. So if you want an unfiltered AI
• 7:24
  model talking to you, go buy a Tesla.
• 7:27
  If you don't want an unfiltered AI talking to you, buy something
• 7:32
  else or get an older Tesla. I think it's 2020.
• 7:36
  Anything with an Intel or an AMD chip can't use grok.
• 7:40
  So anything with those chips on them, make sure that you get one
• 7:45
  of those if you don't want grok in your car.
• 7:48
  So it's a touchy subject for a lot of people.
• 7:51
  Do you want an AI that's kind of rowdy?
• 7:54
  And do you want an AI that will talk back to you and tell you
• 7:58
  horrible things about history? Sometimes it's up to you, You
• 8:02
  know, there's freedom for you to choose.
• 8:05
  So myself personally, I like to test the vehicles, don't really
• 8:09
  want my chatbot to do what Grok has been doing over the last few
• 8:13
  weeks. So it's tough, that's all I can
• 8:17
  say. It's it's a tough situation for
• 8:18
  a lot of people who are in the market to buy a Tesla because if
• 8:22
  you buy 1, you might have to deal with Grok and some sort of
• 8:25
  something that happens and say if you're in your car and your
• 8:30
  family's in the car with you and Grok says something unrestricted
• 8:34
  and you know, a little bit too much or maybe a lot too much, I
• 8:39
  don't know. I don't know what the, what the
• 8:41
  brakes are on this thing. As far as Grok 4 point O in a
• 8:46
  Tesla, maybe it's a like a lesser version of a Grok.
• 8:49
  I'm not sure. They haven't disclosed any
• 8:51
  information about what version it is other than that it's
• 8:54
  version 4. But could it be a dumbed down
• 8:56
  version? Possibly.
• 8:58
  Could it be a version with a lot of brakes on it, a lot of
• 9:01
  fencing? Possibly.
• 9:04
  So just to keep that in mind as you're looking for Teslas,
• 9:09
  because Grok is going to be in every single Tesla going forward
• 9:14
  if it can support it with the chipset that it has now.
• 9:18
  The current debate over Grok and internal whistle blowing could
• 9:21
  push regulatory action, and lawmakers have already raised
• 9:26
  concerns and questions about AI bias and misinformation.
• 9:30
  Should an AI be able to make up history?
• 9:36
  And should an AI be able to sway public opinion by being biased
• 9:41
  one way or the other? They need to do safety audits,
• 9:45
  transparency rules and model reporting standards in order for
• 9:50
  these demands to be met. So for X AI, it's going to erode
• 9:56
  the credibility of its model and raise new legal challenges if it
• 10:00
  continues down this path with Grok.
• 10:03
  And there's pushback now from researchers, which shows that
• 10:07
  it's a very serious issue. And safety lapses have become
• 10:12
  prominent inside one of the most high profile AI companies that's
• 10:16
  also just happens to be owned by Elon Musk.
• 10:21
  Hey, thank you so much for listening today.
• 10:23
  I really do appreciate your support.
• 10:25
  If you could take a second and hit the subscribe or the follow
• 10:28
  button on whatever podcast platform that you're listening
• 10:30
  on right now, I greatly appreciate it.
• 10:33
  It helps out the show tremendously.
• 10:35
  And you'll never miss an episode.
• 10:37
  And each episode is about 10 minutes or less to get you
• 10:40
  caught up quickly. And please, if you want to
• 10:43
  support the show even more, go to patreon.com/stagezero and
• 10:49
  please take care of yourselves and each other and I'll see you
• 10:51
  tomorrow.