Until very recently, it was alt-right people getting frustrated that they couldn't get grok to confirm their delusions. They had tricks to get it to confirm their priors (esp. asking leading questions and demanding a single word response) but they didn't work that well.
When is very recently? I didn't recall any time where Grok wasn't making up answers about how great Elon is and how awful Jewish people, black people, liberals, etc are. It's usually the first test of any model they put out and always gives a ridiculous answer
Recently as in the last few days when it started calling itself "MechaHitler" and scapegoating jewish people after the engineers let Elon ramble for the system prompt.
It was also stating that the life of a single Jew is worth more than that of two million non-Jews.
LLMs can occasionally say crazy stuff, that is not surprising, and I think we should do better than leaning into the outrage machine.
The opposite is how we end up with ridiculous guardrails, like having ChatGPT say that it would rather allow all of humanity to perish than to say the N word, a statement which is orders of magnitude worse, only more publicly palatable.
It was but so were other models before. OP said the twitter to grok feature is a good use case and I agree. Its great for fact checking. For example it will debunk conspiracy theories and misinformation tweets in general. I even asked it about its own hitler meltdown and it rejected its own words (so I must have asked it after they fixed it).
I had the impression, Grok wasn't on Elon's side when it answered my questions or explained tweets.
For a time, yes. Which is why they "fixed it" and it is now calling itself "MechaHitler" and praising Hitler and Musk for "being so based".
That lasted for literal hours before they changed it back. It was clearly just shitposting in a 4chan style way.
10 replies →
While you're not wrong, I feel like they don't make up a significant chunk of @grok's queries. People usually talk about other topics.
This however is a significant chunk of @grok's queries if you only experience it through scrolling Apple News
Until very recently, it was alt-right people getting frustrated that they couldn't get grok to confirm their delusions. They had tricks to get it to confirm their priors (esp. asking leading questions and demanding a single word response) but they didn't work that well.
When is very recently? I didn't recall any time where Grok wasn't making up answers about how great Elon is and how awful Jewish people, black people, liberals, etc are. It's usually the first test of any model they put out and always gives a ridiculous answer
Recently as in the last few days when it started calling itself "MechaHitler" and scapegoating jewish people after the engineers let Elon ramble for the system prompt.
[flagged]
It's not word salad, Grok was literally posting unironic praise for Hitler two days ago.
It was also stating that the life of a single Jew is worth more than that of two million non-Jews.
LLMs can occasionally say crazy stuff, that is not surprising, and I think we should do better than leaning into the outrage machine.
The opposite is how we end up with ridiculous guardrails, like having ChatGPT say that it would rather allow all of humanity to perish than to say the N word, a statement which is orders of magnitude worse, only more publicly palatable.
LLMs can be baited, small changes to system prompts can cause this quite unexpectedly just like many big companies found out by accident.
we fix it and move on.
1 reply →
It was but so were other models before. OP said the twitter to grok feature is a good use case and I agree. Its great for fact checking. For example it will debunk conspiracy theories and misinformation tweets in general. I even asked it about its own hitler meltdown and it rejected its own words (so I must have asked it after they fixed it).