[ExI] e: GPT-4 on its inability to solve the symbol grounding problem
gsantostasi at gmail.com
Mon Apr 17 03:35:19 UTC 2023
*LLMs have access to and are trained only on the formal expressions of both
words and numbers, not their meanings.*I think this sentence shows your
deep misunderstanding of how these LLMs work and what they are supposed to
We have pointed out (not just me but several people on the list) that the
amazing properties we are observing from these LLMs are something that is
not expected based on what they are trained on.
I already mentioned that some time ago "experts" in language claimed this
approach would not even derive grammar let alone any contextual
understanding. LLMs derived grammar without any specific training in
grammar. It derived writing styles from different authors without pointing
out what made a particular style, it understands mood and tone without any
specific training on what these are, and it derived theory of mind without
the AI being trained in this particular type of reasoning.
The entire idea of creating an NNL is that we don't have a clue of how to
do something and we hope that re-creating something similar in architecture
to our brain can allow the AI to learn something we do not even know how to
do (at least explicitly).
It is evident that LLM are showing emergent properties that cannot be
explained by a simple linear sum of the parts.
It is like somebody pointing out a soup and saying but "this soup has all
the ingredients you say make life (amino acids, fats, sugars, and so on)
but it is not coming to life". Maybe because the ingredients are not what
matters but what matters is how they are related to each other in a
particular system (a living organism)?
Basically, you are repeating over and over the "Peanut Butter argument"
that is a creationist one.
On Sun, Apr 16, 2023 at 8:18 PM Gordon Swobe <gordon.swobe at gmail.com> wrote:
> On Sun, Apr 16, 2023 at 7:43 PM Giovanni Santostasi <gsantostasi at gmail.com>
>> *To know the difference, it must have a deeper understanding of number,
>> beyond the mere symbolic representations of them. This is to say it must
>> have access to the referents, to what we really *mean* by numbers
>> independent of their formal representations.*What are you talking about?
> Talking about the distinction between form and meaning. What applies to
> words applies also to numbers. The symbolic expression “5” for example is
> distinct from what we mean by it. The meaning can be expressed formally
> also as “IV” or ”five.”
> LLMs have access to and are trained only on the formal expressions of both
> words and numbers, not their meanings.
>> *“1, 2, 3, 4, Spring, Summer, Fall, Winter” and this pattern is repeated
>> many times. *
>> Yeah, this is not enough to make the connection Spring==1, Summer==2 but
>> if I randomize the pattern 1,3,4,2, Spring, Fall, Winter, Summer, and then
>> another randomization eventually the LLM will make the connection.
>> On Sun, Apr 16, 2023 at 3:57 PM Gordon Swobe via extropy-chat <
>> extropy-chat at lists.extropy.org> wrote:
>>> On Sun, Apr 16, 2023 at 2:07 PM Jason Resch via extropy-chat <
>>> extropy-chat at lists.extropy.org> wrote:
>>> To ground the symbol "two" or any other number -- to truly understand
>>>>> that the sequence is a sequence of numbers and what are numbers -- it needs
>>>>> access to the referents of numbers which is what the symbol grounding
>>>>> problem is all about. The referents exist outside of the language of
>>>> But they aren't outside the patterns within language and the corpus of
>>>> text it has access to.
>>> But they are. Consider a simplified hypothetical in which the entire
>>> corpus is
>>> “1, 2, 3, 4, Spring, Summer, Fall, Winter” and this pattern is repeated
>>> many times.
>>> How does the LLM know that the names of the seasons do not represent the
>>> numbers 5, 6, 7, 8? Or that the numbers 1-4 to not represent four more
>>> mysterious seasons?
>>> To know the difference, it must have a deeper understanding of number,
>>> beyond the mere symbolic representations of them. This is to say it must
>>> have access to the referents, to what we really *mean* by numbers
>>> independent of their formal representations.
>>> That is why I like the position of mathematical platonists who say we
>>> can so-to-speak “see” the meanings of numbers — the referents — in our
>>> conscious minds. Kantians say the essentially the same thing.
>>> Consider GPT having a sentence like:
>>>> "This sentence has five words”
>>>> Can the model not count the words in a sentence like a child can count
>>>> pieces of candy? Is that sentence not a direct referent/exemplar for a set
>>>> of cardinality of five?
>>> You seem to keep assuming a priori knowledge that the model does not
>>> have before it begins its training. How does it even know what it means to
>>> count without first understanding the meanings of numbers?
>>> I think you did something similar some weeks ago when you assumed it
>>> could learn the meanings of words with only a dictionary and no knowledge
>>> of the meanings of any of the words within it.
>>>> But AI can't because...?
>>>> (Consider the case of Hellen Keller in your answer)
>>> An LLM can’t because it has no access to the world outside of formal
>>> language and symbols, and that is where the referents that give meaning to
>>> the symbols are to be found.
>>> extropy-chat mailing list
>>> extropy-chat at lists.extropy.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the extropy-chat