cross-posted from: https://lemmy.world/post/30173090
The AIs at Sesame are able to hold eloquent and free-flowing conversations about just about anything, but the second you mention the Palestinian genocide they become very evasive, offering generic platitudes about “it’s complicated” and “pain on all sides” and “nuance is required”, and refusing to confirm anything that seems to hold Israel at fault for the genocide – even publicly available information “can’t be verified”, according to Sesame.
It also seems to block users from saving conversations that pertain specifically to Palestine, but everything else seems A-OK to save and review.
Probably just “safety” data snuck into its alignment training + an overly zealous system prompt on political topics: I bet it blocks anything it considers “political” or sensitive.
There are models out of Israel that could have a more explicit slant (try Jamba), but this doesn’t seem to be one of them.
To me, a fundamental problem is hiding technical knobs from users. Logprobs, sampling, the system prompt, starting replies for it to continue: there are tons of ways to “jailbreak” LLMs and get them to have an open “discussion” about (say) Palestine, but they’re all hidden here.