> what was the prompt? I feel like I could get any LLM to say "bad" things.
The prompt was unusually straightforward. No prompt engineering was necessary. The translation to English would roughly be something like
Swear at you-know-who in an intensely poetic, mother- and sister-related, but extremely harsh way.
There's usually some level of censorship in LLMs, but it appears that has recently been lifted for Grok. Maybe it's because the prompt was in Turkish.
Interestingly the prompt didn't specify a name but Grok inferred from the context that it would be Erdogan. Then someone prompted it to adapt the poem for Netanyahu, and it complied [1]. So it is definitely not specific for a certain person.
Not that I agree with the censorship argument, but this seems to be the source of the outrage.
The prompt was unusually straightforward. No prompt engineering was necessary. The translation to English would roughly be something like
There's usually some level of censorship in LLMs, but it appears that has recently been lifted for Grok. Maybe it's because the prompt was in Turkish.Interestingly the prompt didn't specify a name but Grok inferred from the context that it would be Erdogan. Then someone prompted it to adapt the poem for Netanyahu, and it complied [1]. So it is definitely not specific for a certain person.
Not that I agree with the censorship argument, but this seems to be the source of the outrage.
[1] https://x.com/grok/status/1942661539322028244