Humanist Discussion Group, Vol. 38, No. 298.
Department of Digital Humanities, University of Cologne
Hosted by DH-Cologne
www.dhhumanist.org
Submit to: humanist@dhhumanist.org
Date: 2025-01-02 18:28:42+00:00
From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 38.285: AI, poetry and readers
Dear Jim,
I took a post December Solstice break, hence the delayed, but
not lessened, thank you for your further response and new
proposal.
Your concise summary nicely outlines, I think, where we agree.
I am, however, still attached to the romantic idea that it
matters, and makes a difference, that human produced text
comes from writing words formed in and by minds. Even if the
mind in question is no longer with us, or never known to us,
and not accessible to us, any attempt to interpret some text
does, I would say, tacitly assume the text is the marks left
by the writing of some words which did, at some point,
originate and reside in a human mind. If we don't assume
this, reading and interpreting and understanding text becomes
like seeing "faces" in the clouds: there are no faces there to
see, not really, it just looks like it. (I'm here ignoring
any belief that The Gods draw pictures in the clouds for us to
see, recognise, and marvel at. Which is close to what I think
is going on when people claim to see words and meanings and
more in today's automatically generated text.)
So, I agree we don't need to know anything about the mind that
originated some words that were then written down, and that we
are left with the text of, to be able to form rational
interpretations of this text. Our interpretation, to be a
fair interpretation, needs no knowledge of what the original
words meant to the mind that created them, and doesn't require
any attempt to surmise this original meaning, nor to be
anything like this original meaning. What we make of the
text, after fair interpretation, is enough. But, I do think
our supposing, or perhaps knowing, that the text we interpret
in our own way was the result of some human mind forming and
writing words, and not some process that just generates text
we can read as if it came from some mind originated words
written down, is what makes forming an interpretation a
sensible thing to do. If we didn't know that, and how, text
results from writing words formed to say things with, I don't
think we'd spend time trying to interpret any text. If that
makes me a romantic, then I'm a romantic.
Your new proposal presents new difficulties for me. This is
based upon the idea, as you state it, that
"... AI generated text is modelled on human minds and
human word-output."
For me, this is a strong claim, and strong claims need strong
reasons and/or evidence to support them. This is because I
hold to a strict definition on what makes something a model of
something: just calling, or describing something as, a model
does not make that something a model of anything, not
necessarily. The term LLM (Large Language Model) is an
example of this kind of empty assertion. In what way is an
LLM a model of language, and is that language in general, or
some language in particular? Are we saying here that because,
given an input sequence of text tokens, the LLM will out put
the text token with the highest estimated probability of being
the next text token in the input sequence, that it is
therefore a model of language? If the answer is yes, it
models what text token comes next in human written text, then
I'm off back to Mars to talk with the Martians. If the answer
is yes, but it models something else about language, I've no
idea what, and nor have any computational linguists I've asked
this question of been able to tell me. What I've got back are
vague explanations along the lines, text is to do with words,
words are to do with language, therefore an LLM is a model of
language. Alice would never allow such lose kind of thinking
through the looking glass.
To have a model of something we must show, well enough, that
is satisfies the Modelling Relation across a sufficient range
of conditions for it to serve as the model we say it is, and
need it to be. What we say is the model must be shown to
stand in a sufficient equivalence relationship with what is
modelled and how this modelled subject is situated in, and
interacts with, its natural or typical situation or context.
This requires making well judged simplifications and
idealisations of what we take to be, or judge to be, important
aspects and behaviours of the subject, and it requires well
defined mappings of chosen [observable] surface features of
the subject to surface features designed to be part of our
model, and either the same, or perhaps different, designed
surface features of the model to be mapped back to what are
chosen to be corresponding surface features of the subject.
And then, we need to verify, validate, calibrate, and test our
model, and present all this for inspection by others, before
we can properly talk of having a model. [This idea of the
Modelling Relation is based upon the work of Robert Rosen,
1985: Anticipatory Systems — Philosophical, Mathematical &
Methodological Foundations, Pergamon Press, and is what I
teach PhDers in a course I do on model making and model using
in research.]
So, we need to ask, I'd say, does a machine built by
[automating the] digging out, from massive amounts of human
made text, huge numbers of detailed statistical relationships
between the mostly unreadable text tokens all the text is
first broken into, model well enough the mental goings on when
a person forms words to say something with, and then writes
these words down? I would say no, it definitely doesn't. The
statistical properties of large amounts of human made text
have no necessary relationship with the mental goings on that
forms and writes words that are meaningful to at least the
author. Alphabets and how we use them to write down words, to
be left with text, are human made artefacts. Certainly useful
artefacts, but they display only the shared ingenuity of human
intelligence. They do not capture, in some understandable
way, the inner workings of human minds, and nor dos any text
written using one of these alphabets. I know of no cognitive
science that claims that a deep statistical analysis of human
made text affords a path into the mental goings on that forms
things to say and words to say them with in the heads of
people who language. As you have consistently pointed out,
there is no needed connection between some text and the
thoughts and goings on in the mind that first formed the words
written down that results in some text, and such a connection
is not needed to make fair interpretation of that text.
Written words do not, you and I insist, deliver into the mind
of readers the meaning the author intends their words to have.
It takes a human reader to have words again from some text --
no human reader, no words -- and it is the mind of the reader
that forms words from the text, and forms, from those words,
what they mean, or might mean, to the reader. And, this
reader made understanding need have nothing to do with what
the words meant to the original author. Using the dug out
statistical patterns of human made text to generate more text
does not, I would say, make this text any kind of model of
human minds, directly nor indirectly, just like the clouds
forming "faces" isn't a model of the cognitive goings on when
a person draws another person's face.
Those "sonnets recently posted" are only sonnets to you who is
able to read the generated text as being the text of written
sonnets. There is nothing in and of the text that makes the
texts sonnets. To have sonnets, and not just the marks on the
page we call text, needs you, someone who knows plenty about
poems and sonnets, and is thus able to read the text as a
sonnet and judge how good is the sonnet you get from your
reading of the text. (It's not the same for me. I am
unqualified to judge how good, or not, this text is as the
text of sonnets.) The machine that automatically generated
this sonnet-looking text is not a sonnet writer, it's just a
text generator that uses statistical patterns found in human
made text to form new text which have the appearance of
sonnets to human readers who know what sonnets are. The text
is 'plastic poems', and, like plastic flowers, no amount of
looking like sonnets makes them real sonnets. Real sonnets
start with some word forming in a human mind, I romantically
want to insist. A not unimportant detail left out of all this
talk about text is the typographical design and formatting of
the text. There are conventional forms used to present the
text of poems that are designed to aid the reading of the text
as poems, and this too, I think matters. But all this detail
is ripped off the texts used to "train" LLMs when it gets
converted into text tokens. It is then put back in to the
output text by some [secrete] internal workings of ChatGPT. It
is not something generated by the LLM, which only generates
text tokens.
On a more detailed point, you say
"... AI can't write anything that I'm aware of that
doesn't have some pre-existing pattern. ..."
First, these automated text generators don't write, they
generate text, but because of the way the so called LLM (Large
Language Model) inside things like ChatGPT is used, new and
never seen before, text patterns are possible; it can, and
therefore might, compose text token patterns into forms that
are not present anywhere in the so called training data. This
is typical of generative mechanisms, statistical or otherwise.
However, it needs you, a poet and human reader of text, to see
that this might be a new sonnet form. ChatGPT, and other
automate text generators don't know what sonnets are, nor what
forms they may take, they only know about text tokens and
loads of statistical relationship between these text tokens
that can be found in loads of human made text.
Only humans write words, machine only generate text. And,
only humans can read text and, by doing so, form words and
thus meanings, in their heads. Machines don't [yet] do this.
ChatGPT does no reading, and does no word forming with which
to say things it has decide to say.
Before I stop, I would like to add here my thanks to,
agreement with, and appreciation of, Willard's call for and
(long time) support of open discussion here, and Miran
Hladnik's and Maurizio Lana's recent public support of, and
arguments for why we need open discussion. Thank you all!
-- Tim
> On 19 Dec 2024, at 10:24, Humanist <humanist@dhhumanist.org> wrote:
>
>
> Humanist Discussion Group, Vol. 38, No. 285.
> Department of Digital Humanities, University of Cologne
> Hosted by DH-Cologne
> www.dhhumanist.org
> Submit to: humanist@dhhumanist.org
>
>
>
>
> Date: 2024-12-18 21:28:37+00:00
> From: James Rovira <jamesrovira@gmail.com>
> Subject: Re: [Humanist] 38.280: AI, poetry and readers
>
> Thanks for the response, Tim. I think I have an idea about how to advance
> our conversation. It may involve a different way of thinking about AI
> generated text.
>
> But first, I was referring to Derrida's 1958 Introduction to Husserl's
> Origins of Geometry. Reading it made me want to write something about the
> importance of the triangle in western philosophy, which goes back to
> Socrates. I can't find the full text online. I might have a .pdf somewhere
> I can send you.
>
> Now, here's the crux of the matter, as I understand it, for you:
>
> "But, [with AI] there was no mind involved in the generation of this text;
> there were no words written down; there was no Shakespeare forming the
> words and writing them down for us to read, and interpret, long after
> Shakespeare's mind is gone."
>
> My previous response to that was that the origin doesn't matter because we
> don't have the mind present. We both agree that a mind originates humanly
> written words, while in the case of AI there is no mind present for that
> specific arrangement of words, but my response was that at the interpretive
> end, mind is equally absent in both cases. That answer wasn't satisfactory
> to you, so you reasserted the difference that an originary, human,
> intentional mind makes for the meaning of words. That is a romantic notion
> to me. In practical terms, the work of interpretation is the same, because
> it is word-based.
>
> But, here's where I'd like to suggest a different idea: words themselves
> are the product of human minds. Patterns of words are the product of human
> minds. AI generates text that follows statistically probable patterns of
> words *in response to a human prompt, and the source text for that
> statistically probable response consists of words** already produced by
> human minds*. AI generated text is modeled on human minds and human
> word-output.
>
> So while the AI itself doesn't have a mind, the AI arrangement of text
> is *indirectly
> *the product of human minds. AI can't write anything that I'm aware of that
> doesn't have some pre-existing pattern. I'm curious what would happen if we
> asked it to invent a new form of poetry? I've innovated a couple short
> poetic forms myself.
>
> Those sonnets recently posted to this list were rather good and could have
> plausibly been written by a human. But AI didn't come up with them out of
> nowhere. I might think the human being who wrote them was a kind of tool,
> but then human beings reduce themselves to tools all of the time. And they
> were "rather good." They weren't great. They weren't self-reflective,
> self-critical, advancing the conversation. They didn't force us to
> reconsider the values being expressed. They were following conventions.
> People do that all of the time. The world is full of minor poets. AI is at
> least a competent poet.
>
> I don't believe I could have this conversation with an AI. It would need
> prompts from me, not responses. So I'm not saying AI and human beings are
> interchangable. I'm only talking about short, discreet, literary products
> such as poems.
>
> Jim R
_______________________________________________
Unsubscribe at: http://dhhumanist.org/Restricted
List posts to: humanist@dhhumanist.org
List info and archives at at: http://dhhumanist.org
Listmember interface at: http://dhhumanist.org/Restricted/
Subscribe at: http://dhhumanist.org/membership_form.php
