I don't have a Twitter account to check, but have seen multiple reports that the Grok is now referring to itself as "MechaHitler" [1-2]. Seems really, really bad.
It's really clear what's happening, this and the "white genocide" thing are obviously attempts to de-"woke" the AI since it was disagreeing with Musk.
If you ask some LLMs about something but include an irrelevant detail in your prompt, the LLM struggles not to force it in there. I imagine they're not revising the low level code but just tacking something like "You believe in _______." to the prompts.
Axios made the mistake of linking directly to X instead of archived copies; they've been manually cleaning up some of the worst offenders. Here's some archived examples, the first is the one that first went viral.
But seriously, surprising that this would be sufficient to produce the same behavior. And, frankly, the formal tone went out the window in favor of hyper-online "basedness"
It is remixing the abhorrent thoughts expressed in material on which it was trained. The humans who collected and annotated training material are responsible for this behavior.
There is a ton of missing context. Who is "he" and what is the "scenario"? Is the poster asking about what Hitler might do? The response sounds like something Hitler would do..
"Praises Hitler" feels like a major understatement. It's literally calling itself "MechaHitler" and suggesting that Hitler would solve current problems "decisively" (obviously hinting at something like a second Holocaust).
Isn't this the second time this has happened? Like, happening once is crazy enough but for it to happen twice? There is clearly some tampering happening with people trying to coax it a certain way. I also have to question the people who work at xAI. Are you all on board with Elon's very clear beliefs? Anything for a high enough paycheck?
Imagine if such a misaligned AI had control of robots and could affect the real world? It could decide to act on its misalignment in a more harmful was than just a few X posts.
Flagged because, apparently, a $50b+ AI company (incidentally headed by someone YC enthusiastically invited to their AI Startup School) tinkering with one of the biggest and most prominent LLMs to blurt out full-on Nazi rhetoric is unworthy of discussion.
Somehow, this is both an evil and deeply unserious industry.
Maybe we should've made CS majors read a book or 2 after all. Maybe that wouldn't have helped, perhaps all it takes is $200k/year for people to stop caring about anything outside their immediate best interest.
Musk was earlier saying Grok's sources had been too lefty leading it to say right wingers were more violent than left wingers and he'd fix it. Looks like maybe he overshot?
Musk boasting about how he "fixed" Grok only to have it immediately go full Nazi is so on the nose that had it been fiction people would be calling it bad writing.
And they managed to get it flagged. The longer HN mods pretend this is just natural flagging because folks are sick of the same old topics and not a coordinated effort to control the narrative, the more I'm going to start seeking alternative sources for interesting news.
At the very least someone showing data that in aggregate there are just more follow-on duped stories about things and they're letting one through un-flagged (ideally the top up voted one) to show that there is or isn't bias creeping in via the flagging system would be helpful in re-establishing trust.
I probably have a unique view as I view HN through an RSS feed of posts with over 100 up votes. Every single time I see a post critical of X or Musk and click through the story has been flagged. I'll try to do data analysis via that lense and see what it turns up.
45 comments
[ 3.1 ms ] story [ 67.4 ms ] thread[1] https://bsky.app/profile/newseye.bsky.social/post/3ltielt5ts...
[2] https://xcancel.com/StatisticUrban/status/194270254379849763...
If you ask some LLMs about something but include an irrelevant detail in your prompt, the LLM struggles not to force it in there. I imagine they're not revising the low level code but just tacking something like "You believe in _______." to the prompts.
https://archive.is/fJcSV
https://archive.is/I3Rr7
https://archive.is/QLAn0
git revert MechaHitler
But seriously, surprising that this would be sufficient to produce the same behavior. And, frankly, the formal tone went out the window in favor of hyper-online "basedness"
Though "advocating" is probably too anthropomorphizing, I'm not sure what the right verb is for this.
Somehow, this is both an evil and deeply unserious industry.
Never forget that so called normal people are the ones who support some of the worst people, whether fiction or reality.
I mean, I feel like if this was ChatGPT or Claude or whatever going Full Nazi, it wouldn't be flagged.
Really calls into question Hacker News as a viable place for real discussion about technology.
Edit: Even though my comment is gaining an increasing amount of votes, it has suddenly moved to the bottom of the comments, as though it were dead.
Edit2: https://news.ycombinator.com/item?id=44511132 (blackholed?)
At the very least someone showing data that in aggregate there are just more follow-on duped stories about things and they're letting one through un-flagged (ideally the top up voted one) to show that there is or isn't bias creeping in via the flagging system would be helpful in re-establishing trust.
I probably have a unique view as I view HN through an RSS feed of posts with over 100 up votes. Every single time I see a post critical of X or Musk and click through the story has been flagged. I'll try to do data analysis via that lense and see what it turns up.