[ExI] Emily M. Bender — Language Models and Linguistics (video interview)

Gordon Swobe gordon.swobe at gmail.com
Mon Mar 27 05:12:05 UTC 2023


On Sun, Mar 26, 2023 at 10:39 PM Jason Resch via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

>
>> It knows how the words "left" and "right" relate to each other and to
>> other symbols related to spatial dimensions, syntactically but with no
>> access to the referents to give them meaning.
>>
> The fact that  GPT can construct a coherent essay from what is to it
>> meaningless symbols is to me no less amazing than the fact that it can
>> create a mathematical model from meaningless symbols. It's all basically
>> the same amazing process, a giant web of inter-relationships between
>> meaningless symbols that have no meaning to it, but which do have meaning
>> to us on account of it was trained on the forms and patterns of our
>> language.
>>
>
> But your (and I think Bender's) assertion was that there was no possible
> way to learn any meaning whatsoever without a Rosetta stone, or without
> programming in some model of reality from the start.
>

It must be getting late where you are.:) I meant that there is no Rosetta
Stone *for GPT* to understand the meanings of the symbols. We need no
Rosetta Stone if we are using the English version of GPT.

If that's true, how then does the LLM come to learn the spatial meaning of
> a word like 'down' when all the model encounters are "meaningless symbols"
> which are themselves only defined in terms of other "meaningless symbols"
> ad infinitum?
>

It never learns those meanings, but because it understands the grammatical
(syntactic) relationships  between the symbols, it can nonetheless
construct strings of symbols that have meaning to us. As I've pointed out,
some of those strings are potentially dangerous, which is why the models
have humans to steer them. Older versions of ChatGPT used to say all sorts
of crazy things. ChatGPT is not a person.

-gts

>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230326/94c8a0af/attachment.htm>


More information about the extropy-chat mailing list