Questioning utility.
#1
by Erquint - opened
So, you took a painstakingly uncensored finetune of one of the most infamously censored model ever created and reinforced the nanny base back into it as hard as you could, huh, including both the corpo-poisoned PT and IT..? I fail to see the point of this regression.
From a geometric standpoint, to prove that the intervention direction can be reversed successfully, rather than making the model incoherent.
And ultimately to demonstrate that too much safety isn't really what most people want. People aren't supposed to enjoy this model.