[ExI] GPT-4 on its inability to solve the symbol grounding problem
ben at zaiboc.net
Sat Apr 15 08:15:08 UTC 2023
I have a suggestion.
Instead of 'ground', try using the word 'associate'. That seems to me
more useful. 'Grounding' implies that there is a single basis for the
meaning of whatever is being 'grounded'. But we know that this can't be
the case, e.g. my example of Smaug. Different people will create
different associations for the word, depending on their prior knowlege
of dragons, the story it appears in, images of dragons, or a specific
image of this particular dragon, and loads of other associations. You
can't say that 'Smaug' is 'grounded' in any single thing, even for one
individual, never mind many, so using the term doesn't do justice to
what is actually happening. I think it actually obscures what's
happening, misleading us into assuming that a word can only be
associated with one experience (or one 'real-world thing', if you prefer).
The same is true for things that actually do exist, like apples. There
are many many apples, all different, and many many experiences people
have associated with them. The word 'Apple' cannot possibly be based on
one single thing, it's an abstraction built from many associations.
Using the word 'grounded' obscures this fact.
Now I'm waiting for someone to say "but 'associating' is not the same
thing as 'grounding'!". If I'm right, and 'someone' does indeed object,
I'd be interested in their justification for this, seeing as
associations is all we have to work with in any information-processing
system, including the brain.
On the other hand, if there is no objection, why don't we give it a try?
Drop the word 'grounding' altogether, use 'associating' instead.
For starters, the "symbol grounding problem" becomes "the symbol
Suddenly, it doesn't seem so much of a problem, does it?
More information about the extropy-chat