While of course we don't know exactly how this happened, it seems like a heck of a coincidence that Elon Musk has recently managed to sell this debunked conspiracy theory to Donald Trump.
Comments
Log in with your Bluesky account to leave a comment
Dont forget, in trying to get your message out there, use of acronym without first using the longer form, Large Language Model, can impede the casual "scroller"
Remember when Musk wanted to buy Wikipedia but couldn't?
This is the same playbook.
(Just a reminder as we try to figure this out: you can't ask an LLM why it did something and expect a reliable answer. So the posts where Grok says Musk programmed this have to be considered unreliable for now)
That’s true and a valid point. But it’s also likely that it’s challenging to tweak model weights alone to bring up something as uncommon as “white genocide.”
Anthropic showed it was technically possible to identify and amplify classes/tokens (see: making model think it was the Golden Gate Bridge). But White Genocide is a hard topic to see arising organically. My money would be on it being true that Grok has additional prompt instructions.
Since the system prompt is also a prompt like any other on the LLM chat history, it can reference it to produce a response to questions about it's inner workings.
But these things aren't self-aware: it's just generating statistically plausible text based on token weights and chat history.
There is a wrinkle that is being missed I think. Social media is a black box. We don’t know if the vast majority of people we interact with are legitimate in any way. Or how what we are being shown or not shown is decided.
This exchange shows an alarming feature - Grok is ingesting material from X quickly enough to power viral misinformation. Here it's citing a thread posted on X by this account less than 5 hours before (reposted by @whstancil.bsky.social). (Note: Grok is hallucinating the timestamp.) [repost - typo]
Comments
https://www.theguardian.com/technology/2025/may/18/musks-ai-bot-grok-blames-its-holocaust-scepticism-on-programming-error
Almost like there's an underlying theme here...
So long as LLMs are black boxes in the control of billionaires, we can't let LLMs become trusted information sources.
But agree entirely
This is the same playbook.
(Just a reminder as we try to figure this out: you can't ask an LLM why it did something and expect a reliable answer. So the posts where Grok says Musk programmed this have to be considered unreliable for now)
But these things aren't self-aware: it's just generating statistically plausible text based on token weights and chat history.