[ExI] e: GPT-4 on its inability to solve the symbol grounding problem

[Ben Zaiboc]

On 19/04/2023 15:11, Gordon Swobe wrote:
>
> Ben, I can't locate the message, but you asked my thoughts on the 
> difference between a language model solving what you have called the 
> word association problem and its solving the symbol grounding 
> problem. In my view, the difference lies in the fact that 
> understanding statistical associations between words does not require 
> knowledge of their meanings. While this distinction might not make a 
> practical difference, it becomes important if the question is whether 
> the model genuinely understands the content of its inputs and outputs 
> or merely simulates that understanding.


What are you telling us Gordon??!!

Exams are useless! Oh. My. Dog.

All those exams!!

I'm going to need therapy.

All those exams completely failed to determine if my understanding was 
real or just simulated!

Watch out, guys, the Degree Police are coming for us!

Surely, Gordon, there must be some test to tell us if our understanding 
is real or simulated?

Oh, wait, you said "this distinction might not make a practical 
difference". Might not? Well, we should pray to our canine slavemaster 
in the sky that it doesn't!

Because, to be honest, I kind of suspect that /all/ my understanding, of 
everything, is merely simulated. In fact, I think even my perception of 
the colour Red might be simulated.

I might as well turn myself in right now. <Sob>.


Aaaanyway, having got that out of my system, I do believe you've twisted 
my words somewhat, and I wasn't referring specifically to LLMs, but 
information-processing systems in general, and particularly human 
brains. I was trying to point out that the 'symbol grounding problem' is 
solved by considering the associations between different models and 
processes in such systems, and you even agreed with me that when people 
use 'referents' they are using the internal models of things, not 
referring directly to the outside world (which is impossible, I don't 
remember if you explicitly agreed to this as well, but I think so). 
Therefore 'symbol grounding' = associating internal models with 
linguistic tokens. I said I don't know how LLMs work, or whether they 
use such internal models.

I also pointed out that these models can be constructed from any signals 
that have consistent associations with sensory inputs, and could be the 
result of any process that inputs data (including text).

Now it may be that 'understanding' does require these internal models, 
and it may be that LMMs don't have them. As I said, I don't know, and am 
not making any claims about either thing. So, just for the record, I'm 
not one of these 'Zealots' you seem to have constructed an internal 
model of (remember what I said: just because you have a model of 
something, that thing doesn't have to actually be real).

In my view, you are correct that "understanding statistical associations 
between words does not require knowledge of their meanings". That's 
hardly a controversial position. But that's not to say that 
understanding statistical associations between words cannot /lead/ to 
knowledge of their meanings. Several people have already given you 
several examples of how it can.

My little ramble above deals with the difference between genuinely 
understanding something and merely simulating the understanding.

(I think we should also be on our guard against systems simulating 
addition, as opposed to genuinely adding, not to mention a few other 
things).

Ben
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230419/2b3d9e39/attachment.htm>