[ExI] Bender's Octopus (re: LLMs like ChatGPT)

[Giovanni Santostasi]

Everything is simulated. Our entire mental life is simulated. The brain is
making up the world from limited information and create models all the
time. We would not be able not just understand but even see or experience
anything if we didn't SIMULATE the world around us. We do it all the time.
We confabulate all the time, like NLMs exactly do.
They work because they do EXACTLY what we do.


On Fri, Mar 24, 2023 at 12:19 AM Gordon Swobe <gordon.swobe at gmail.com>
wrote:

> You wrote to GPT: A friend of mine says NLM or any other AI cannot
> understand anything without concrete referents. I disagree, which arguments
> could you give to support my position?
> That is not actually what I am saying. It makes no difference whether the
> referent is "concrete" or not.
>
> As GPT says in supposed defense of your position:
>
> They can "simulate" understanding by leveraging their knowledge of how
> certain words or concepts are used in different contexts.
> That's right. Language models can *simulate* understanding. It is not real
> understanding. It is simulated understanding. They mimic what understanding
> looks like objectively.
>
> I remember writing at length on this list about 15 years ago about how the
> digital simulation of a thing is not the same as the thing simulated.
>
> -gts
>
>
>
> On Fri, Mar 24, 2023 at 12:57 AM Giovanni Santostasi <
> gsantostasi at gmail.com> wrote:
>
>> A friend of mine says NLM or any other AI cannot understand anything
>> without concrete referents. I disagree, which arguments could you give to
>> support my position? While it is true that NLMs and other AI models do
>> not have direct experiences or access to concrete referents like humans do,
>> they can still demonstrate a certain level of understanding of abstract
>> concepts. Here are some arguments to support your position:
>>
>>    1.
>>
>>    Pattern recognition and statistical relationships: NLMs, like
>>    ChatGPT, are trained on vast amounts of text data, and they learn by
>>    recognizing patterns and statistical relationships within that data. They
>>    can generate responses that demonstrate an understanding of abstract
>>    concepts based on these patterns, even without direct experience of
>>    concrete referents.
>>    2.
>>
>>    Contextual understanding: Through their training, NLMs develop an
>>    ability to understand context and respond appropriately to various
>>    situations. They can "simulate" understanding by leveraging their knowledge
>>    of how certain words or concepts are used in different contexts.
>>    3.
>>
>>    Flexibility and adaptability: NLMs can adapt to new information and
>>    generate responses that take into account various factors, including the
>>    context, prior knowledge, and the specific question or statement they are
>>    responding to. This adaptability allows them to address abstract concepts
>>    and ideas, even without direct access to concrete referents.
>>    4.
>>
>>    Success in various tasks: NLMs have shown success in a range of
>>    natural language processing tasks, such as question-answering, translation,
>>    and summarization, which involve understanding and manipulation of abstract
>>    concepts. Their ability to perform well on these tasks is evidence of their
>>    capacity to work with abstract ideas without direct access to concrete
>>    referents.
>>
>>
>> On Thu, Mar 23, 2023 at 11:44 PM Gordon Swobe <gordon.swobe at gmail.com>
>> wrote:
>>
>>> I can already hear someone saying "but we will include photographs of
>>> objects in the training so they have referents," but this still does not do
>>> the trick. These digital photographs can be displayed to the human operator
>>> of the chatbot, but the bot itself sees only 1s and 0s, on's and off's. It
>>> can detect colors by wavelength, but still this is only digital data. It
>>> does not see the colors. Likewise with shapes. It is turtles (one's and
>>> zero's) all the way down with no referents.
>>>
>>> -gts
>>>
>>> On Fri, Mar 24, 2023 at 12:18 AM Gordon Swobe <gordon.swobe at gmail.com>
>>> wrote:
>>>
>>>> Nobody least of all me questions that GPT-4 will be capable of amazing
>>>> feats, and that eventually these language models will surpass humans in
>>>> terms of what we can call intelligence or what I might for sake of clarity
>>>> prefer to call apparent intelligence. The question here is whether they
>>>> will know what they are saying given that they are trained only on the
>>>> forms of words with no access to the meanings or referents.
>>>>
>>>> Adrian has made the excellent point a couple of times that this is like
>>>> the first contact problem in science fiction, and actually like the first
>>>> contact problem between any two cultures with completely different
>>>> languages. Q: When Kirk and Spock beam down to a new planet with
>>>> intelligent alien life, how will they learn to communicate? A: With
>>>> referents.
>>>>
>>>> Spock will point to himself and say "Spock." Kirk will point to himself
>>>> and say "Kirk." Kirk will point to a rock and say "rock." Kirk and Spock
>>>> use these kinds referents to initiate  communication. If our alien friend
>>>> wants to communicate, he will point to the rock and "fwerasa" (or whatever
>>>> is his word for rock). He will point to himself and say his name, and so
>>>> on. Eventually, Spock and the alien will learn how to translate a few
>>>> words, and from there the process of understanding begins.
>>>>
>>>> Now, what if they don't beam down to the planet and listen to only
>>>> digital radio signals coming from the planet and send digital radio signals
>>>> in return? No communication is possible as there are no referents. It's all
>>>> noise.
>>>>
>>>> -gts
>>>>
>>>>>
>>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230324/218b9da6/attachment.htm>