Humanist Discussion Group

Humanist Archives: Feb. 19, 2022, 6:23 a.m. Humanist 35.541 - from GPT-3 to helpful agents

				
              Humanist Discussion Group, Vol. 35, No. 541.
        Department of Digital Humanities, University of Cologne
                      Hosted by DH-Cologne
                       www.dhhumanist.org
                Submit to: humanist@dhhumanist.org




        Date: 2022-02-18 17:02:50+00:00
        From: Simpson, Erik <SIMPSONE@Grinnell.EDU>
        Subject: Re: [Humanist] 35.539: from GPT-3 to helpful agents

Dear Willard,

Your example of a writer trying to get unstuck remind me of Vauhini Vara's
account, from August of last year, of using GPT-3 to get unstuck on another kind
of writing project. The piece's subtitle is "I didn't know how to write about my
sister's death--so I had AI do it for me." As you will see in the piece, "do it
for me" is an oversimplification. In fact, the process is, in its way,
conversational. Or if it is not conversational, I am not sure I have a word for
what happens.

You can read it at https://believermag.com/ghosts/. I wonder what you will
think.

Best wishes,

Erik


On 2/18/22, 2:07 AM, "Humanist" <humanist@dhhumanist.org> wrote:

Date: 2022-02-18 07:52:25+00:00
From: Willard McCarty <willard.mccarty@mccarty.org.uk>
Subject: an artificial interlocutor

Mark Wolff's remarks on GPT-3 have led me to think in more specific
terms anout the kind of artificially intelligent agent I'd like to have.
Comments on the following would be welcome.

Let's say the situation is that of an academic who is stuck in the midst
of writing a research paper. This person has done all the necessary
research, read all the material needed, but is having trouble coming up
with an argument or continuing one. Let's say that recourse to a trusted
friend to talk through the project is impossible. The question is, how
could an artificially intelligent agent help?

The desired agent is nothing like a Mechanical Turk, who puts on an
impressive performance of some kind. The academic does not know what
questions to ask, so an agent that plays tricks with whatever he or she
might ask would not work. Training data such as is available online in
however massive quantities seems unlikely to be helpful, and the kind
that would be--published scholarly literature--isn't freely available in
sufficient quantities, so it seems best to assume that the task is to
draw on what the writer already 'knows' but doesn't know that he or she
knows. How about total access to the subject's hard disc (ignoring for
the moment privacy issues). Hence the agent has to be quasi-psychiatric,
but which style of psychiatry depends on how the agent and the writer
work together. Carl Rogers' therapeutic method (hence the 'Rogerian'
of ELIZA's DOCTOR) is certainly attractive, but I imagine that there are
other candidates.

What would be the medium?

Words? The agent could be somewhat like a child who won't give up asking
'why?', but may well come back with 'do you mean...?' or 'if X then Y?'
Perhaps we'd give the agent access to all the textual material on the
writer's computer as training data. Or perhaps we'd train the agent on
all the folklore of the writer's culture.

Images? Perhaps, as with Hermann Rorschach's famous ink blots, images of
the right sort would trigger thoughts. In The Glass Plate Game that I
mentioned earlier drawings on cards are used. But the objective here is
not psychoanalytic. What sort of images might work?

Music?

Comments please.

Yours,
WM


--
Willard McCarty,
Professor emeritus, King's College London;
Editor, Interdisciplinary Science Reviews;  Humanist


_______________________________________________
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