Humanist Discussion Group, Vol. 39, No. 96.
Department of Digital Humanities, University of Cologne
Hosted by DH-Cologne
www.dhhumanist.org
Submit to: humanist@dhhumanist.org
[1] From: scholar-at-large@bell.net <scholar-at-large@bell.net>
Subject: modalities and sequences (79)
[2] From: Gabriel Egan <mail@gabrielegan.com>
Subject: Re: [Humanist] 39.94: repetition vs intelligence: on LLMs (203)
[3] From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 39.91: repetition vs intelligence: on LLMs (110)
--[1]------------------------------------------------------------------------
Date: 2025-07-23 20:49:28+00:00
From: scholar-at-large@bell.net <scholar-at-large@bell.net>
Subject: modalities and sequences
Willard
A little orthogonal contribution to the repetition vs intelligence thread…
I hold that at a certain level of abstraction the audible and the visible
sensory modalities operate in a similar fashion:
[quote]
5.27
The human senses, whatever their number and relations, produce events. Events
can be connected. This production of events can be experienced, can be induced,
can be guided. Memory plays a major role in this process. Attention can be
alternatively devoted to percept and to the act of perception. The
possibilities for metacommentary are connected to the possibilities for memory.
Cognitively this allows humans to preserve the trace of something happening at a
certain time. Events connected in a series of episodes lead to narratives. The
transformation of discrete somatic signals into sequences begins to explain
cross-modal encoding.
[/quote]
"Storing and Sorting"
https://lachance.artsci.utoronto.ca/S6.HTM
<https://lachance.artsci.utoronto.ca/S6.HTM>
I suggest that sensitivity to the translatability of sensory modalities (a
hesitancy to adopt dichotomous relations between hearing and sight) accompanies
a machine-view or structuralist approach to human culture. The above quoted
continues:
[quote]
5.28
Although not dealing with sequences, Alexander Alland drawing upon the work of
Charles Laughlin and Eugene d'Aquili, Biogenetic Structuralism, suggests that
anatomical and physiological factors enhancing cross-modal association are
responsible for the emergence of conceptualization ("Roots of Art" 13-14). ;
Developing an anthropology of art, Alland posits an aesthetic-cognitive function
for which he offers the term transformation-representation. His notion is
allied to narrative or sequence processing. He argues:
Art is an emotionally charged and culturally central storage device for complex
sets of conscious and unconscious information. Structure guards information in
well-ordered and easily retrievable forms. It also allows for a certain amount
of variation (transformation) without loss of total information or organization.
Transformation is something that is likely to occur by accident, but it is also
likely to be part of the aesthetic game in which playing with form is a major
element. Transformation without significant changes in over-all structure keeps
the game exciting at the same time as essential information is guarded.
(Artistic Animal 41).
As form is to storage and circulation, sequence is to narrative and narration.
[/quote]
Preceding this formulation was a consideration of sensoria as not only receiving
but also sending:
[quote]
5.25
Conceiving story as storage and story as algorithm is the key to imagining a
sensorium that is more than merely receptive, to theorizing
one that is interactive in regards to its modalities and its environment. The
stumbling block in imagining such a sensorium has been proper
theorizing of the means of translating from one modality to another. Verbal
language seemed to be the best candidate. However it privileged
sight and hearing, the distance senses, over those of closer contact: smell,
touch, and taste.
5.26
In re-evaluating the closer contact senses, especially their action under
conditions of distress or extreme pleasure, one discovers that the
sensorium not only is a receiver but also a dispatcher of information. The
senses are not only receptors. The senses also transmit. By their
operation the senses provide events for interpretation. The blinking of eyes,
the cocking of an ear, the flicker of a tongue, all signal.
[/quote]
Thank you for your kind indulgence and apologies for the lapidary prose,
François Lachance
Scholar-at-large
--[2]------------------------------------------------------------------------
Date: 2025-07-23 20:49:01+00:00
From: Gabriel Egan <mail@gabrielegan.com>
Subject: Re: [Humanist] 39.94: repetition vs intelligence: on LLMs
Dear Humanists
Tim Smithers wrote:
> The claim that I say representation
> system building needs perfection is your
> claim [that is, Gabriel Egan's], not
> mine. I didn't use this word, nor
> in any way imply perfection is needed,
> as a reading of even just what you quote
> from my long post shows.
The following is a snippet of the
phrases that I think are where
Smithers not only implies but baldly
asserts that perfection is needed
in a representational system:
> . . . in all cases without exception ...
> . . . as designed, only as designed,
> . . . can only happen as designed, and
> always happen . . . in all the conditions
> and situations . . . all, without fail . . .
(Smithers, Humanist, 12 June 2025)
Since we won't agree on this, I leave it
to the judgement of Humanists whether
using "all ... without exception ... only
... only ... always ... all ... all ...
without fail" constitutes a counsel of
perfection.
Smithers also wrote:
> Please show us what it is I repeatedly
> said wrong about what Large Language
> Models actually do . . .
Sure. Here's an example:
> . . . if, as you [Egan] claim here,
> each dimension of this vector
> space somehow encodes meaning,
> then, to be a dimension of a
> vector space, each dimension
> must encode a unique meaning, and
> a meaning that is orthogonal
> to all other meanings on all
> the other dimensions. This
> orthogonality is a necessary
> property of the dimensions of
> a vector space. Perhaps you'd
> like to show us how this is true,
> and true for every one of the
> dimensions, without exception,
> say, for the 12,888 dimensions
> of the vector embedding space
> used in GPT-3 . . . . You'll
> need to tell us what the 12,888
> unique and orthogonal meanings
> are.
(Smithers, Humanist, 10 June 2025)
Let's keep using Smithers's "12,888
dimensions" for GPT-3 although I
believe he actually meant 12,288.
To keep on board those unfamiliar
with some of the terms, I will
gloss the ones I think not all
Humanists will be familiar with.
Those who know this stuff please
bear with me.
The feature called 'orthogonality'
is the characteristic of two
lines or directions being at
an angle of 90% degrees from
each other. When we move from
a dimensionality of one, as in
the one-dimensional number
line we first learn at school,
to a dimensionality of two,
as in an x/y plot, we necessarily
put the second axis (the y axis)
at 90 degrees from the first
(the x axis, our original number
line).
How many vectors (= lines with
direction and length) can you
draw in an x/y plot so that for
every possible pair of lines
you've drawn the lines in that
pair are at 90 degrees to each
other?
Answer: exactly two.
If you start with an x/y plot and
draw on it two lines ('a' and 'b')
that are orthogonal (at 90 degrees)
to each other, you cannot add a
third line ('c') to the drawing
and have it be orthogonal to both
'a' and 'b'. That is what I mean
above by "so that for every possible
pair", as in a-and-b, b-and-c, and
a-and-c, the lines are orthogonal.
The only way to add a line 'c'
that is orthogonal to 'a' and
orthogonal to 'b' is to invoke
a third dimensional 'z' that
comes directly up off the plane
of our paper x/y drawing to hover
above it in space. Now we can add
a third line 'c', rising directly
off the page, that is orthogonal
to 'a' and orthogonal to 'b'.
In three-dimensional space, the
number of vectors we can draw
so that each pair of vectors
is orthogonal is 3. And so on
for four-dimensional space
(4 vectors) and upwards.
In n-dimensional space, the number
of vectors you can draw so that
every pair is orthogonal is 'n'
itself. This might seem obvious
to some, since it is really just
how we define our notion of
dimensionality.
In Smithers's GPT-3 example, the list
of numbers assigned to each word is
12,888 numbers long. Each long list
of numbers can be said to define a
point in 12,888-dimensional space.
If each dimension encodes a meaning,
it should be clear why we want the
dimensions to be orthogonal. If one
dimension represents 'blondness'
and another represents 'sugariness'
then we want to be sure that a vector
that encodes 'a bit more sugariness'
(used to mark, say, the difference
between fruit juice and ketchup)
should in principle encode nothing
about increased or decreased blondness.
The two concepts are unrelated, so
that our saying that something is a
bit more sugary than something else
should not entail us saying that it
is also a bit more blond too.
We can see why Smithers is sceptical
that 12,888 dimensions is enough
for each dimension to hold a concept
in human meaning. It just doesn't
seem like enough dimensions for
there to be one for each meaning
in human existence. Thus he asks me
to "tell us what the 12,888 unique
and orthogonal meanings are".
Here is where Smithers's counsel of
perfection leads him astray. We don't
need to maintain perfect orthogonality
between our vectors that encode
particular meanings. If we demand perfect
orthogonality the number of vectors
we can have in n-dimensional space so
that every pair is orthogonal is just
'n'. But if we go from perfect
orthogonality to, say, demanding that
the angle between each pair of vectors
lies between 88 and 92 degrees (so,
approximately orthogonal) then the
number of vectors we can put in n-
dimensional space such that every
vector pair is approximately orthogonal
rises exponentially as 'n' rises.
In 12,888-dimensional space we can
draw a lot more than 12,888 vectors
such that every pair is nearly
orthogonal. This gives us a lot more
'space' for particular directions to
encode meanings. The exponentiality
that underlies this doesn't have much
effect until the number of dimensions
gets large, which might be why LLMs
had an unexpected boost in performance
once investigators started using a
lot more dimensions than previous
investigators had.
Regards
Gabriel Egan
--[3]------------------------------------------------------------------------
Date: 2025-07-23 08:01:51+00:00
From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 39.91: repetition vs intelligence: on LLMs
Gabriel,
You say, of what I wrote ...
"At root his objection is that, on principle, a machine
cannot wield meanings, which is what humans do. Behind all
such insistence that humans do something that machines
cannot do there lies, I suspect, human hubris and fear of
obsolescence."
No! I didn't present any such objection. None of what I have
written in any recent posts has been about what machines or
humans can and cannot do. Why would it be? That's not what
we're talking about. We are talking about how systems like
ChatGPT, and the LLMs they use, really work, and what they
really do and don't do, and the misleading understanding of
all this we get from the fairy stories told about it all.
You haven't told us what you mean by meaning, so there's no
way I would try to assert "... a machine cannot wield
meanings." And, until you tell us what you mean by "wield
meanings" here, I also cannot say if this is something humans
do. So I wouldn't try to. If, for you, to "wield meanings"
is somehow like to wield an axe, I may say no, humans don't do
this. But, poetry is, it seems to me, to be appreciated, or
not, as poetry, and not presented as a technical explanation
for how a sophisticated machine works or doesn't.
I do think it may be possible to build machines that
understands human language, and which can therefore fairly be
described as dealing in meanings, or, at least understand
limited but useful amounts of human language. Some twenty
years ago Urko Esnaola (then a PhDer) and me designed and
built a language understanding system for a robot we used.
This was for a simple artificial language which used musical
notes for the sounds the words were made of: it was a "spoken"
language, but it also had a written form. This artificial
language, of course, had nothing remotely like the
sophistication of human languages. It didn't need to have.
But the robot did understand what we said to it in this
language, which we did by whistling the notes of the words, or
by playing them on a simple xylophone, or a recorder, or
violin, and we understood what the robot said to us in this
language. Importantly, the robot also heard and understand
what it said, and it used this to keep tract of the
interaction, build expectations of what may come next, and
adapt it's sound generation to the different acoustics of the
building it operated in: it was a real building where other
people worked, not a lab. This artificial languaging worked
reliably, and over distances of tens of meters, and in
situations with plenty of other background noise, and
sometimes lots of people. Kids delighted in it. It was
implemented using a Minsky "Society of Mind" type of system,
with a high quality omni-directional microphone, an ordinary
loud speaker, used only small amounts of cpu and memory, and,
of course, ran in real time. This was far from human level
languaging, but it does show that at least simple machine
languaging is possible.
ChatGPT, with it's massive LLM inside it, doesn't do anything
like this; it doesn't understand human language; it does no
language processing. ChatGPT, as I have explained at some
length, deals in text tokens, only text tokens, and it uses
very large amounts of detailed statistics of patterns of these
text tokens found in gigantic amounts of human written text,
to automatically generate text formed using these text tokens.
Yes, it generates grammatically correct text. Yes, we can
read this text as being about something, often something we
expect the text to be about, in some interaction with ChatGPT.
But this doesn't mean, therefore, ChatGPT generates text by
writing about what we read the text to be about. It doesn't.
It can't write; it only generates text tokens. It also can't
read; it only transforms text into text token sequences, and
then builds a vector for each token (using an ambiguous
similarity measure) which is said to be similar to the vectors
for other tokens it has a high point wise mutual information
value with: vectors which cannot be shown to satisfy a
Representation Relation for word meaning no matter how you
specify word meaning, which never is publicly specified by
LLM builders. ChatGPT doesn't know, understand, or reason
about any of what the text it generates appears, to us, to be
about: it has no mechanisms for knowledge representation, for
reasoning with represented knowledge, or for presenting
explanations. It has mechanisms for presenting text made from
its text tokens, but always sugar coated to make it look like
the text has been written by ChatGPT in a confident, never
qualified, first person style: all deliberate deception.
With so much statistics of text token patterns found in so
much human written text, we should expect to be able to build
machines which can, using these statistics, automatically
generate text that appears to be about something when we read
it. This is a property of most human written text: when we
read and understand it, it appears to be about something, and
we take it that this something was intended by the author. So
we'd expect the statistics of human written text, turned into
sequences of text tokens, to reflect this property in some
statistical correlations. An LLM has enormous amounts of this
statistics of text token patterns found in human written text,
but it's nothing more than statistics of text token patterns.
And these statistics, as with the statistics of any data, is
only at the level of the data, of the text tokens in this
case. No amount of fairy story telling can change this;
there's no magic that turns text token statistics into meaning
representations, no matter how much statistics you have.
ChatGPT is like a "talking" parrot which appears to say things
we understand. No matter what it sounds like to us, this
parrot isn't doing any languaging. Or, do you think it is?
-- Tim
_______________________________________________
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
