[ExI] Is the GPT-3 statistical language model conscious?

Bill Hibbard hibbard at wisc.edu
Sat Oct 10 22:46:27 UTC 2020

GPT-3 is trained solely from text data, so it's all
about how text relates to other text. In human brains
text also relates to sensory and motor experience. So
the word "chair" relates to our physical experiences
with chairs.

The training text for GPT-3 may cover just about
everything that can be said about chairs, so GPT-3 may
do a pretty convincing job talking about chairs. But
the lack of connection with the physical world will
severely limit GPT-3. Kind of reminds me of the old
definition of a consultant: knows 100 ways to make
love but doesn't know any women.

More information about the extropy-chat mailing list