I once saw a list of instructions being passed around that were intended to be tacked on to any prompt: e.g. “don’t speculate, don’t estimate, don’t fill in knowledge gaps”
But you’d think it would make more sense to add that into the weights rather than putting it in your prompt and hoping it works. As it stands, it sometimes feels like making a wish on the monkey paw and trying to close a bunch of unfortunate cursed loopholes.
Adding it into the weights would be quite hard, as you would need many examples of text where someone is not sure about something. Humans do not often publish work that have a lot of that in it, so the training data does not have examples of it.
I once saw a list of instructions being passed around that were intended to be tacked on to any prompt: e.g. “don’t speculate, don’t estimate, don’t fill in knowledge gaps”
But you’d think it would make more sense to add that into the weights rather than putting it in your prompt and hoping it works. As it stands, it sometimes feels like making a wish on the monkey paw and trying to close a bunch of unfortunate cursed loopholes.
Adding it into the weights would be quite hard, as you would need many examples of text where someone is not sure about something. Humans do not often publish work that have a lot of that in it, so the training data does not have examples of it.
Simple solution: don’t use the stupid things. They’re a waste of energy, water and time in the best case.