Humanist Discussion Group, Vol. 36, No. 375.
Department of Digital Humanities, University of Cologne
Hosted by DH-Cologne
www.dhhumanist.org
Submit to: humanist@dhhumanist.org
Date: 2023-02-02 20:04:22+00:00
From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 36.369: ChatGPT as co-author
Dear Jim and Norman,
First, Jim, David Bowie explains his use of "cut-up" in this
short BBC News item.
Cut up technique- David Bowie
<https://youtu.be/m1InCrzGIPU>
But, to answer your question about how the program he used
worked, this short BBC News item helps, I think: the program
was 'loaded up' with texts chosen by Bowie, which it then
'randomly' cut-up into pieces which Bowie then decided how to
make use of.
How David Bowie used 'cut ups' to create lyrics - BBC News
<https://youtu.be/6nlW4EbxTD8>
But, I've not found details of this program: what it was
written in; how it did the random cutting; how the texts where
input -- by typing them in, I suppose? ... If anybody has
anything on this it'd be fund to know more.
Second, Norman, Perplexity is a kind of diversity index used
to estimate the uncertainty of probability distributions, see,
for example, Perplexity on Wikipedia
<https://en.wikipedia.org/wiki/Perplexity>.
Applied to Large Language Model (LLM) stuff, here are two
technical explanations which probably present all you need to
know, and perhaps more.
Perplexity of language models revisited
From basic information theory to practical computation
By Pirmin Lemberger
<https://towardsdatascience.com/perplexity-of-language-models-
revisited-6b9b4cf46792>
and
Perplexity of fixed-length models
at Huggin Face
<https://huggingface.co/docs/transformers/perplexity#perplexity-of-
fixedlength-models>
I used GPTZero to test the following chunk of what I got back
from ChatGPT (in my second interaction).
ChatGPT:
AI models such as ChatGPT do not have the capacity for
self-awareness, agency, or personal interests, and
therefore, cannot make decisions about authorship. As AI
models, they are tools created and controlled by humans, and
their use in generating text is guided by human input and
control."
with the result from GPTZero:
"Your text is likely to be written entirely by a human"
Similarly testing other chunks from the same interaction
results in GPTZero indicating plenty of AI generated text.
I tested this particular set of words because I wondered if
they might bear the marks of some human intervention: has
someone at OpenAI arranged for this kind of response to be
made? The probability of this sequences of words (tokens) in
the ChatGPT training data must be tiny, if not effectively
zero, it seems to me, given it only exists because of ChatGPT
and its ilk: I don't think anybody, or anything, made text
like this before we got ChatGPT.
It is also my impression, from playing with GPT-3 and ChatGPT,
that the texts they generate about AI and AI generated text
has become quickly more coherent, and defensive of a
human-sided point of view: no self-awareness, agency, or
personal interests, etc.
So, perhaps not everything found in ChatGPT output is only AI
generated. The plastic text is being sugar coated, sometimes.
Best regards,
Tim
> On 1 Feb 2023, at 08:53, Humanist <humanist@dhhumanist.org> wrote:
>
>
> Humanist Discussion Group, Vol. 36, No. 369.
> Department of Digital Humanities, University of Cologne
> Hosted by DH-Cologne
> www.dhhumanist.org
> Submit to: humanist@dhhumanist.org
>
>
> [1] From: James Rovira <jamesrovira@gmail.com>
> Subject: Re: [Humanist] 36.366: ChatGPT as co-author (54)
>
> [2] From: Norman Gray <norman.gray@glasgow.ac.uk>
> Subject: Re: [Humanist] 36.366: ChatGPT as co-author (42)
>
_______________________________________________
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
