[ot][crazy][spam] Transformer Models: Finding Rationality Without It
This paper was on my queue: https://arxiv.org/pdf/2203.14465.pdf It does more than it simply says it does. It applies a number of techniques that are relatively obvious, and have been roughly ignored, to the situation. It is sad it uses a neural language model to perform such a simple task as rational reasoning. But large neural networks seem to be what I'm looking at nowadays. And they have flexibility.
Like the language models in question, I have trouble with rationality. It's very hard for me. In the paper, they found a research program to force their language model into, where the model was rewired to become rational. This required forcing it to behave in many normal skills, such as writing out its steps before performing them. Things AI used to do already, in past decades.
participants (1)
-
Undiscussed Horrific Abuse, One Victim of Many