[ExI] Another ChatGPT session on qualia

Giovanni Santostasi gsantostasi at gmail.com
Thu Apr 27 00:45:12 UTC 2023


*As Tara pointed out so eloquently in another thread, children ground the
symbols, sometimes literally putting objects into their mouths to better
understand them. This is of course true of conscious people generally. As
adults we do not put things in our mouths to understand them, but as
conscious beings with subjective experience, we ground symbols/words with
experience. This can be subjective experience of external objects, or of
inner thoughts and feelings.Pure language models have no access to
subjective experience and so can only generate symbols from symbols with no
understanding or grounding of any or them. I could argue the same is true
of multi-model models, but I see no point to it is as so many here believe
that even pure language models can somehow access the referents from which
words derive their meanings, i.e, that LLMs can somehow ground symbols even
with no sensory apparatus whatsoever.*

All this is just based on ideology and not careful thinking. It is clear to
me now.
But let's reply in a logical fashion.
1) What is one of the mostly common first words for a child? Moma. But Moma
doesn't refer to anything initially for a child. It is a babbling sound
child make because some programming in our brain makes us test making
sounds randomly to train our vocal cords and the coordination between many
anatomical parts that support vocal communication. But somehow the babbling
is associated with the mother. Who is doing the grounding? Mostly the
mother, not the child. The mother overreacts to these first babbling
thinking that he is calling her and self assign this name to herself, which
is basically the opposite of grounding a specific sound to a specific
intended target, lol. It is mostly in the mother's head. Then the mother
teaches the child this is her name and the child learns to associate that
sound with the mother. This is such a universal phenomenon that in most
languages the name for mom is basically the same. This alone should destroy
any simplistic idea that humans learn language or meaning by making a 1 to
1 association with some real object in the physical world. It is much more
complex than that and it has many layers of interaction and abstraction
both at the individual and at the social level.
2) When the mother (notice again even in this case we are talking about a
complex interaction between mother and child) points to an object and says
APPLE and the child listen to the mother what exactly is going on there? If
Gordon was right that there is some grounding process going on there, at
leas his very naive understanding of grounding, the association will happen
more or less immediately. It doesn't, the mother has to show the apple
several times and repeat the name. But then finally it happens the child
repeats the name. That repetition doesn't mean the child made the
association, it could simply mean it simply repeats the sound the mother
makes. In fact, that is an important step in learning a language first the
child behaves like a little parrot (being a parrot actually is a good thing
to learn languages not bad as Bender seems to claim). The true
understanding of the word apples most of the time comes later (there are
situations where the mother will point to the apple, make the sound and the
child doesn't respond until one day he hold an apple and says apple) when
the child sees an apple or holds an apple or tastes an apple and says
"APPLE". Is this grounding as Gordon understands it?
NO ! Why? Well the mother pointed not at one single apple in this process
but many. If it was grounding as naively understood then it would have
confused the child more and more to point to different objects and them
being called apples. These objects didn't have the same exact size, they
maybe had different colors (some red, some yellow), and slightly different
tastes, some more sour some more sweet. They are different. So I don't say
that what Gordon calls "grounding" is actually the opposite of grounding to
be contrarian but because I deeply believe this idea of grounding is
bullshit, utter bullshit and in fact it is the core of all our
misunderstanding and the fact most of current linguistic doesn't understand
language at a higher level that is necessary to understand languages not
just in humans but the alien minds of AI.
This process cannot be grounding as 1 to 1 one directional association
between the object and the meaning of the object.
For the child to make the connection it requires understanding what the
mother means by pointing to the object and uttering a sound (the 2 are
connected somehow that is not a simple idea to process), that the mother
doesn't mean this particular object in front of me at this particular time,
that a red apple and a yellow apple can be still apples (so the child needs
to figure out what they have in common and what they don't and what is not
important to identify them as apples), the child needs to understand that
if the apple is cut in slices, it is still an apple and so on and on and
on. Do you see how bullshit the idea of grounding is?
How a cut apple (just thought about this) can be still an apple? But the
child somehow knows !
It is not the grounding that counts here in learning the language but the
high-level abstraction of associating a sound with an object, the fact that
different objects can be put in a broad category, that objects can be cut
in pieces and be still together as a whole or in part (half an apple is
still an apple) the same object, not physically but conceptually and from
an abstract point of view.
There is no grounding without all this process of abstraction and this
process of abstraction is in a way "GOING AWAY FROM GROUNDING", in the
sense that it requires literally moving away from the specific sensory
experience of this particular object in front of me. The grounding is at
most a feedback loop from abstraction to object, from object to
abstraction, and so on. It is not at all the main component in giving
meaning to language. It is easy to see how one can build a language that is
all abstractions and categorization. We have shown this many times when we
showed that we can build a symbolic language made of 0 and 1s or how we can
build math from the empty set and so on. But what I have discussed above
shows that abstraction comes before grounding and it is necessary for
grounding to happen.
The phenomenon of grounding is really a misnomer.
What happens in this exercise of naming things is that it allows us to see
connections between things. The objects is not what is important but the
connections, the patterns. Now in the case of the mother teaching a
language to the child that has to do with objects in the real world, it
happens that this language has a survival value because learning patterns
and regularities in the natural world, being able to think about them,
being able to communicate to others about these patterns ("A wolf is coming
!) has an evolutionary advantage so yes, it has an additional value, it is
not useless.
But the fact that most human language has some relevance to understanding
the physical world doesn't show AT ALL that the association with the
physical world is required for giving meaning to a language.
I don't know how to make this argument more clear and compelling.
One could write an entire book on this and maybe even invent an entire
language that has nothing to do with real physical objects and it is all
self-referential. It is obvious to me the brain did that (anything the
brain knows is electrical train spikes anyway, including sensory
experience) and that LLMs did that too.
But it is clear from my arguments above that Gordon and the linguist are
wrong.

By the way, I pointed out that Umberto Eco, that was one of the most
renowned semiotics experts had a similar understanding of the process of
grounding and call it the "reference fallacy". For him, a sign (that is
what words are) only points to another sign in a never-ending process. The
never-ending is not necessary for most communication because at a point we
simply decide we think we know enough about what something means (we use
basically Bayesian inference in our brains to do that) and LLMs do the same
settling on some probabilistic value of the meaning of the words it uses.
If something is highly probable probably is true (pun intended).

Giovanni














On Wed, Apr 26, 2023 at 3:19 PM Adrian Tymes via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> On Wed, Apr 26, 2023 at 3:05 PM Gordon Swobe <gordon.swobe at gmail.com>
> wrote:
>
>> On Wed, Apr 26, 2023 at 3:45 PM Adrian Tymes via extropy-chat <
>> extropy-chat at lists.extropy.org> wrote:
>>
>>> On Wed, Apr 26, 2023 at 2:33 PM Gordon Swobe via extropy-chat <
>>> extropy-chat at lists.extropy.org> wrote:
>>>
>>>> This is the section of GPTs' reply that I wish everyone here understood:
>>>>
>>>> > My responses are generated based on patterns in the text and data
>>>> that I have been trained on, and I do not have the ability to truly
>>>> > understand the meaning of the words I generate. While I am able to
>>>> generate text that appears to be intelligent and coherent, it is
>>>> > important to remember that I do not have true consciousness or
>>>> subjective experiences.
>>>>
>>>> GPT has no true understanding of the words it generates. It is designed
>>>> only to generate words and sentences and paragraphs that we, the end-users,
>>>> will find meaningful.
>>>>
>>>> *We, the end-users*, assign meaning to the words. Some
>>>> people mistakenly project their own mental processes onto the language
>>>> model and conclude that it understands the meanings.
>>>>
>>>
>>> How is this substantially different from a child learning to speak from
>>> the training data of those around the child?  It's not pre-programmed:
>>> those surrounded by English speakers learn English; those surrounded by
>>> Chinese speakers learn Chinese
>>>
>>
>> As Tara pointed out so eloquently in another thread, children ground the
>> symbols, sometimes literally putting objects into their mouths to better
>> understand them. This is of course true of conscious people generally. As
>> adults we do not put things in our mouths to understand them, but as
>> conscious beings with subjective experience, we ground symbols/words with
>> experience. This can be subjective experience of external objects, or of
>> inner thoughts and feelings.
>>
>> Pure language models have no access to subjective experience and so can
>> only generate symbols from symbols with no understanding or grounding of
>> any or them. I could argue the same is true of multi-model models, but I
>> see no point to it is as so many here believe that even pure language
>> models can somehow access the referents from which words derive their
>> meanings, i.e, that LLMs can somehow ground symbols even with no sensory
>> apparatus whatsoever.
>>
>
> Agreed, for the record, but I figured the point needed clarifying.
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230426/c9126b8b/attachment.htm>


More information about the extropy-chat mailing list