Re:ChatGPT Political Compass
by nonrandomstring on Thursday December 29, 2022 @09:31PM
(#34181077)
> That's sort of like where a parent will give a wrong, but easy
answer to a credulous child when asked a question, and a different
one to adults with more discernment.
Somewhat, yes. But I think it better fits the description somewhere
between a sociopath and Harry G. Frankfurt's definition of a
bullshitter, who has utter disregard for the truth and will say
anything in the moment that deals with a question qua problem. It's
essentially defensive and uses language instrumentally only in that
pursuit.
The problem here is not what LLM-AI is capable of, but what
capabilities people ascribe to it. The danger comes from what we
believe or project/personify onto it. It has no essential concept of
truth. But we want it to.
Re:ChatGPT Political Compass
by narrator on Thursday December 29, 2022 @10:03PM
(#34181312)
On the politicized stuff they use InstructGPT[1] to make sure it gives the "right" answer. This is human in the loop reinforcement learning. The human training it was probably instructed to give it a pass on any answer that was not "toxic."
Somewhat, yes. But I think it better fits the description somewhere between a sociopath and Harry G. Frankfurt's definition of a bullshitter, who has utter disregard for the truth and will say anything in the moment that deals with a question qua problem. It's essentially defensive and uses language instrumentally only in that pursuit.
The problem here is not what LLM-AI is capable of, but what capabilities people ascribe to it. The danger comes from what we believe or project/personify onto it. It has no essential concept of truth. But we want it to.