The last paragraph is interesting.
Sure, generating harmful responses 7.5% of the time is better than 51% of the time. But we’re still far from a safe LLM, if that’s even possible.
I’d rather companies would NOT make LLMs available publicly as a service unless the rate is < 1%. What they’re doing now isn’t responsible.
What they’re doing now isn’t responsible.
Would it be responsible if every response would start with “I lie x % of the time but here’s my response:”?
Would it be responsible to sell canned beef which makes you sick 7.5% of the time?
What if there was notice saying “Only 7.5% of our delicious canned beef contain listeria”?
This is how to cover your ass. This is not how to be responsible.
“snubs”?
I have no problem idea what that means. Even with the context.
In this case it means opts for alternative
