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

[Giovanni Santostasi]

This has been done before:
Can you find a study that deals with the idea of building a language
without physical referents?

While I cannot browse the internet for the latest studies, I can point you
to an interesting experiment that deals with the idea of creating a
language without direct physical referents. This experiment is known as the
"Talking Heads Experiment," conducted by Luc Steels and his colleagues in
the late 1990s.

The Talking Heads Experiment involved two robotic agents with cameras that
were programmed to "learn" a shared language by pointing at objects in
their environment and communicating with each other about what they were
seeing. The robots were not provided with any predefined language or
vocabulary. Instead, they had to develop their own communication system by
interacting with each other and learning from their experiences.

Through this process, the robots were able to create a shared lexicon and
grammar for describing objects and their properties, such as color and
shape. This experiment demonstrated that it is possible for agents to
develop a language without direct physical referents, relying on
interaction and feedback to build a shared understanding of their
environment.

Although the Talking Heads Experiment involved robots with cameras that
allowed them to perceive their environment, the experiment still
illustrates the potential for creating a language based on interaction and
learning, rather than relying on direct physical referents.

For more information on the Talking Heads Experiment, you can refer to the
following sources:

   1. Steels, L. (1999). The Talking Heads Experiment: Origins of words and
   meanings. Laboratorium, Antwerpen.
   2. Steels, L. (1998). The origins of syntax in visually grounded robotic
   agents. Artificial Intelligence, 103(1-2), 133-156.


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/20230324/32d92ef9/attachment.htm>