Latest / Elon Musk Podcast / xAI Cuts 500 Grok Trainers as Musk Bets on Automation

Transcript

xAI Cuts 500 Grok Trainers as Musk Bets on Automation · 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, the
• 0:08
  shape, SpaceX, Tesla X, the Boring Company and Neurolink.
• 0:13
  I'm your host Will Walden. XAI laid off nearly its entire
• 0:19
  data annotation team last week, cutting around 500 contractors
• 0:24
  who had been responsible for training Grok, which is the
• 0:27
  chatbot that is used on X slash Twitter.
• 0:32
  And these workers made-up the bulk of the workforce behind the
• 0:36
  scenes. They labeled data the curated
• 0:39
  prompts for Grok and fine-tuned Grok's responses across a wide
• 0:45
  range of tasks. Now, what does XAI plan on
• 0:50
  doing? Are they going to replace these
• 0:52
  people with some other people, or are they going to allow Grok
• 0:56
  to moderate itself now? X AI didn't publicly acknowledge
• 1:00
  the layoffs, though workers received notice without advance
• 1:03
  warning, and most of them were contractors based outside the US
• 1:07
  working through third party staffing firms.
• 1:10
  The team had been assembled over the past year to accelerate
• 1:13
  Grok's progress from an early stage model into a commercial
• 1:17
  project. Now cutting them so suddenly
• 1:21
  signals a shift in strategy that X AI hasn't explained to anybody
• 1:26
  yet. And the timing matters for this
• 1:28
  because X AI only launched Grok 2 in March and has already begun
• 1:31
  teasing more versions of Grok in the future.
• 1:35
  The company claims that Grok 3 is training on a custom compute
• 1:39
  cluster with 100,000 NVIDIA H1 hundreds in.
• 1:43
  Musk has promised Grok will surpass opening eyes ChatGPT
• 1:47
  four and five, and will also match GPT 6 when it comes out.
• 1:52
  And there they've limited 500 trainers during this transition.
• 1:58
  And it forces the question of how Grok will be improved
• 2:01
  without humans during this feedback at this kind of scale.
• 2:08
  And these weren't just generic contractors, though many of them
• 2:10
  worked full time on language tasks tailored specifically to
• 2:13
  Grok's rebellious tone and X based content integration.
• 2:17
  They created prompts, reviewed output, and flagged errors.
• 2:21
  Some were building specialized instruction databases to help
• 2:24
  Grok, answer coding questions, generate memes, or hold long
• 2:28
  form conversations. Several had backgrounds in
• 2:31
  machine learning or linguistics, even if they weren't formally
• 2:35
  engineers. And the layoffs came abruptly.
• 2:38
  One worker described being locked out of the system mid
• 2:40
  task, while others found Slack access revoked before receiving
• 2:44
  any formal notice. Most had been expecting long
• 2:47
  term contracts and had been told their roles were important to
• 2:50
  Grok's future intentions. Sounds very corporate to me and
• 2:54
  it's happened to a bunch of people in the past.
• 2:57
  Different corporations. If this has happened to you,
• 2:59
  please leave a comment on your podcast platform right now.
• 3:03
  It's happened to me. I worked in e-commerce before.
• 3:06
  One day I tried to log into the system so I could check to see
• 3:09
  if we had any orders coming in early in the morning and if so,
• 3:14
  I wanted to check out there. If I was doing analytics for the
• 3:19
  company, I was running the whole e-commerce system there, the
• 3:22
  whole the whole site, and I was checking analytics to see what
• 3:28
  these people were doing early in the morning.
• 3:30
  I would wake up at 5:30 or 6 and check the analytics and I
• 3:34
  couldn't log into my e-mail. I couldn't log into the
• 3:37
  e-commerce site. I couldn't log into our employer
• 3:41
  portal or our employee Porter portal.
• 3:44
  So I called my boss at 6:00 in the morning and asked him what's
• 3:48
  going on? Like could we get this fixed?
• 3:49
  I don't know if it's a glitch or whatever.
• 3:51
  He said yeah, we shut it all down.
• 3:54
  So unlike Xai, he shut the whole e-commerce site down but also
• 3:58
  didn't tell me what was going on because he told me, and this is
• 4:04
  pretty scummy, but he told me I had no right to know what was
• 4:07
  going on. Even though I had just invested
• 4:10
  a year of my life into building this platform from scratch and
• 4:14
  we were getting good sales, we were profitable and I don't know
• 4:18
  why they shut it down. To be honest with you.
• 4:20
  I have no idea what the heck this guy was thinking because it
• 4:23
  was going to be a successful business and luckily I landed on
• 4:26
  my feet. Here I am doing this podcast and
• 4:28
  I'm successful at being a contractor and also helping
• 4:32
  people with their businesses and their YouTube and podcast
• 4:35
  businesses. So it landed on my feet.
• 4:38
  All is good on my end. But these people, 500 people
• 4:43
  were laid off. I'm not going to use the word
• 4:46
  laid off. That's too nice.
• 4:50
  Elon fired them, left them without paychecks, left them
• 4:54
  without a way to earn money for their families, left them in
• 4:59
  limbo. Who knows if these people were
• 5:01
  living paycheck to paycheck, we don't know what they were being
• 5:03
  paid. Maybe them being laid off now
• 5:08
  they're, you know, they have no money for food for their family.
• 5:11
  Maybe they have 3 or 4 kids and none of those kids can eat now
• 5:14
  because Elon Musk fired them. I'm being harsh, but it's the
• 5:18
  harsh reality of corporate America and corporate Elon.
• 5:22
  He just gets rid of people whenever he feels like it.
• 5:24
  And I'm a news guy. I'm going to tell how how it is.
• 5:27
  I like the companies that Elon is part of.
• 5:31
  If that makes you frustrated and mad that I said that, think
• 5:35
  about it from a perspective, a broader perspective, not just in
• 5:39
  your head, not just what you feel, but think about it like
• 5:43
  this. If that happened to you and you
• 5:45
  had four kids, I think you'd be upset.
• 5:50
  I think one kid even. It's just you.
• 5:51
  What are you going to do now? Hopefully you have some savings.
• 5:54
  You're eating into your savings because nobody told you what's
• 5:56
  going on. Nobody told you that after X
• 5:59
  amount of time, you're going to need a new job.
• 6:02
  So please, you know, make the right choices, Save some money
• 6:06
  because after this job, you're going to need to find another
• 6:09
  one. And with this job market, good
• 6:11
  luck, especially in tech right now, good luck.
• 6:14
  AI is taking all your jobs and XAI and Grok, well, maybe
• 6:20
  they're being trained by a Grok version in the background now.
• 6:25
  Musk has also hinted at plans to automate more of the data
• 6:27
  labeling and training process using synthetic data, but XAI
• 6:32
  has not explained what it's using to replace human
• 6:34
  supervision of this. Maybe they have some higher ups
• 6:38
  out of those 500 people. Maybe they have their managers
• 6:40
  or something. I'm not sure to keep track of
• 6:43
  the AI, and additional addition to that, we don't know what's
• 6:48
  going on behind the scenes. In traditional AI development,
• 6:50
  human annotators play a critical role in model alignment in
• 6:54
  response tuning. Scaling grok with minimal
• 6:56
  oversight creates real risk, especially for a product
• 6:59
  designed to answer questions sourced from a volatile platform
• 7:02
  like X. There's a huge amount of insane
• 7:07
  people on X, and we all know that.
• 7:09
  And if they're training their data on X, good luck, you know,
• 7:16
  sourcing that. But you know, X and Grok and AI
• 7:22
  models are getting so advanced, they might just have the answer
• 7:26
  now. This isn't the first time that
• 7:27
  Musk has gutted a support team in pursuit of automation.
• 7:31
  After taking over Twitter, he eliminated most of the
• 7:33
  platform's content moderation staff and replace them with AI
• 7:37
  driven filtering and community based reporting.
• 7:40
  That move introduced ongoing accuracy and abuse problems that
• 7:44
  continue to affect the platform up until today.
• 7:48
  Now, the Grok layoffs look like a similar gamble with the same
• 7:51
  potential consequences. We're going to have to keep an
• 7:54
  eye on Grok's next release and see how wild it gets.
• 7:58
  Musk has said that Grok will eventually be a personal AI
• 8:01
  assistant fully integrated into X and Tesla platforms, and that
• 8:05
  vision requires tight performance, adaptive
• 8:08
  conversations in a contextual awareness across different user
• 8:11
  inputs. And removing the people who
• 8:14
  trained Grok to handle that complexity calls into question
• 8:17
  how far along that vision actually is, or whether XAI is
• 8:22
  just skipping critical steps to meet aggressive release targets
• 8:26
  and get more investor money. Some of the laid off workers had
• 8:29
  only recently been on boarded and we're in the middle of
• 8:32
  building new data sets. They just got there.
• 8:35
  Others have been asking for clearer task guidelines and
• 8:38
  internal documentation to keep up with shifting model behavior.
• 8:41
  Instead of getting answers, they just got fired.
• 8:44
  This suggests that XAI may not have had a cohesive plan for
• 8:48
  ongoing model improvement, even as it is prepared for the next
• 8:52
  version of Grok. Now, XAI has repeatedly
• 8:57
  positioned itself as an engineering 1st operation.
• 9:01
  Now, that posture often devalues humans that that are required to
• 9:06
  build useful and aligned AI systems.
• 9:09
  But removing the people who are responsible for shaping how Grok
• 9:13
  actually talks to people, actually responds to people and
• 9:18
  reasons, strips the model of an essential feedback loop, one
• 9:22
  that can't be easily replaced with more compute alone.
• 9:26
  The data annotation team function as Grok's immune
• 9:29
  system, catching and correcting errors before they reached end
• 9:33
  users. Their absence will show up in
• 9:35
  the product itself. If Grok starts sounding more
• 9:39
  erratic, a little crazy, missing context, or, as AI gurus like to
• 9:46
  say, hallucinating or just straight up lying to you more
• 9:51
  confidently, this is where it all started.
• 9:55
  The 500 people that were fired without reason because they
• 10:03
  think they can do it without them.
• 10:05
  Hey, thank you so much for listening today.
• 10:07
  I really do appreciate your support.
• 10:09
  If you could take a second and hit this subscribe or the follow
• 10:12
  button on whatever podcast platform that you're listening
• 10:14
  on right now, I greatly appreciate it.
• 10:17
  It helps out the show tremendously and you'll never
• 10:19
  miss an episode. And each episode is about 10
• 10:22
  minutes or less to get you caught up quickly.
• 10:25
  And please, if you want to support the show even more, go
• 10:28
  to Atreoncom Stage Zero. And please take care of
• 10:33
  yourselves and each other. And I'll see you tomorrow.