Humanist Discussion Group, Vol. 38, No. 367.
Department of Digital Humanities, University of Cologne
Hosted by DH-Cologne
www.dhhumanist.org
Submit to: humanist@dhhumanist.org
[1] From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 38.312: AI, poetry and readers: Calvino, neuroscience & intention (482)
[2] From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 38.316: AI, poetry and readers (112)
--[1]------------------------------------------------------------------------
Date: 2025-02-19 17:50:05+00:00
From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 38.312: AI, poetry and readers: Calvino, neuroscience & intention
Rewinding the tape 1
Back in early January I said I would respond to your message,
Gabriel, of 2025.01.08. With PhD teaching now done for a few
weeks, here's my reply, with apologies for the long wait.
You say ...
"... things don't have to be built the same to work the
same. Aeroplanes work like birds in how they fly: they
generate lift by deflecting a moving airflow over
specially shaped wings. Two things don't have to be
physically identical to be functionally similar (that is
"like" each other, in my phrasing) The lenses in my
spectacles work like the lenses in my eyes, for
instance."
Aeroplanes do indeed fly by generating lift the same way bird
flight generate lift (but not propulsion forces) -- I
originally trained as an aeronautical engineer. So, in this
case, we can say, show that, and well explain how, the "lift"
force generated in both cases is the same kind of force
generated by the same aerodynamic phenomenon. It's not just
similar, it's identical, and this identity relationship
matters here. The lift generated by a hot air balloon, for
example, is a force too, but a force not generated in the same
was as the lift force generated by wings moving though the
air.
Similarly, as we currently understand the optics of light,
which, I think we will easily agree, is, today, a good and
reliable understanding, does allows us to say that, show that,
and explain well that, the way the lenses of your eyes work is
the same way the lenses in your spectacles work, optically:
they both refract light in the same way.
But, both these examples only work as different ways of
achieving an identical effect -- aerodynamic lift, and bending
of light waves -- because we have shown that they are the same
force generated the same way, and both are real light wave
bending by refraction in a transparent medium. Until we can
do this, any claims of being the same, of being identical, are
speculation, and possibly mistaken. Claims of identity are
always strong claims, and strong claims require strong
support: strong evidence, argument, and, best, the same
theory.
The claim that brains use neural networks and computers
(programmed in a certain way) also use neural networks, and
are thus different implementations of the same kind of
machine, is a strong identity claim, but one that nobody has
shown to be true. If, Gabriel, you think this identity has
been well enough demonstrated, please show us the evidence and
arguments and theory for this. In my view, also shared by
many others I know, just calling the things we program in
computers "neurones" does not, and cannot, make what we
implement the same as real neurones. It does not make them
real neurones implemented in a different way. Calling them
both neurones does not, and cannot, make them the same thing.
Nouning maketh not the nounded!
You say, that when I say the so called neural networks of
computer based systems are not the same as the neural networks
we find, and know some things about, in brains, begs the
question of "what is a neural network." No, of course it
doesn't. We know, albeit to a currently limited extent, what
real neural networks are, the networks of real neurones we
find making up [in very large numbers] the networks of
neurones in real brains. There is no dispute about this. We
do know what real neurones and real neural networks are!
There is no question begging here.
Just because we do not yet know all there is to know and
understand about these real neurones and the real networks
they go together to form in real brains, does not mean we do
not know what real neural networks are, and that we can
therefore properly say other [very large] networks of simple
input summing thresholded logical switching devices,
implemented using digital computation, are the same,
identically the same. This is just wishful thinking at best,
the same as the dogma Connectionist AI still rests upon, or
it0s deliberate deception.
Real neurones, found in real brains and naturally occurring
nervous systems, as we currently understand them, come in a
variety of different types, and go together to form what look
like, under a microscope, and appear to function as, different
(sub)systems, when looked at using certain real-time brain
scanning techniques. But, one thing all these different types
of neurones have in common, is that they are all kinds of
basic timers in the network systems they form a part of. And,
it is this intrinsic temporal dynamics of their individual
behaviour that is basic to the functioning of the networks
they are a part of. Neurones are understood and studied as
dynamical timing devices, and have been for a long time in the
neurosciences, using mathematical formulations and models of
this dynamics, as we understand it. See, for example, F C
Hoppensteadt "An introduction to the mathematics of neurones"
(1986), a book recommended to me by some neuroscientists who
worked on numerical models of real neural networks that were
understood to provide the basic control mechanism of the six
legged walking in Stick Insects. A dynamical phenomenon
needing just the right kind of timing. Or, if you prefer more
on the real stuff, look at Valentino Braitenberg, "On the
Texture of Brains" (1977), and chapter 4 in particular, which
explains five different types of signal transmission
mechanisms between neurones, making it hard to see them as
just input summing thresholded two state logical switch
devices, and hard to think the dynamics of these different
signalling mechanisms is of no importance in understanding what
real neurones do and how networks of these function.
This is important because it means our most basic
understanding of real neurones is as dynamical in-time
devices, and not as logical switching devices which can be
fully understood as, and computationally implemented as
over-time two state switching devices. Everything that
happens takes time to happen, but for some things that happen,
how long it takes to happen makes no difference to what does
happen, whereas, for other things that happen, how long it
takes to happen matters completely. We distinguish these two
kinds of happenings by calling them either over-time
happenings -- when how long it takes to happen doesn't matter
-- or calling them in-time happenings -- when how long it
takes to happen does matter.
In-time devices and over-time devices are thus two distinct
classes of devices, and using over-time devices to model
systems built of in-time devices will not, and cannot, capture
important, and basic, aspects of the dynamical behaviour of
the in-time devices and thus of the dynamics of the systems
built using them, not without applying further in-time
conditions, constraints, and other, extra, mechanisms.
Digital computation is intrinsically an over-time kind of
happening: it always takes some time to happen, but it makes
no difference to the computation that happens how long it
takes to happen. Of course, it usually matter to us how long
the computation takes, but this only makes our use of digital
computation and in-time like use. The utility, to us, of some
computation is greatly affected by how long it takes for the
computation to happen, but no matter how long it does take, it
is always the same computation. It doesn't make the
computation we use a kind of in-time happening; it is always
an over-time happening. It is only when we apply further
dynamical conditions and constraints, and extra mechanisms, as
we do is so called real-time systems, that the computations
are made to have in-time behaviour, but these additional
constraints, conditions, and mechanisms, are not computational
in nature. In other words Turing Machines, and the useful
approximations to these we build and use, and call computers,
are over-time machines and not in-time machines. Brains, I
submit, are, on the basis of all we currently know and
understand about what they consist of, how they function, and
what they do, are in-time machines, and the basic units they
are built from -- all the different kinds of neurones -- are
themselves in-time devices.
Now, as you mention, Gabriel, early Connectionists did use
analogue electronic devices to build simple models of real
neurones, and [very] small neural network models, but these
were soon displaced by Rosenblatt's work on the Perceptron,
first using his Mark 1 Perceptron, a special purpose machine
which used digital computation, and then "simulated" on the
soon to arrive, and more convenient to use, digital computers.
Curiously Rosenblatt's book, "Principles of Neurodynamics"
(1962), despite it's title, doesn't deal with the intrinsic
dynamics of real neurones. It adopts the characterisation
established and promoted by Warren McCulloch and Walter Pitts
in their publication "A logical calculus of the ideas immanent
in nervous activity" (1943), of neurones as logical over-time
logical two state switching devices. This is probably because
at the time, in the late 1930s and early 1940s, the intrinsic
dynamical, and thus necessarily in-time, behaviour and
functioning of real neurones and neural networks, was only
poorly known and understood. And, I would say, the
convenience afforded by McCulloch and Pitts characterisation
of neurones as simple input summing thresholded logical
switching devices, which can be easily implemented using
digital computation, meant that Connectionists have seldom
gone back to either question of revise this convenient, but
simplistic, characterisation of what real neurones are, what
they do, and how they do it.
So, what I'm saying here is that what are called by
Connectionists, neurones, or artificial neurones, are a
different class of device, that this difference matters in
fundamental ways, that this all makes real neurones and real
neural networks, found in brains, quite different from the so
called neural networks people say they have built in their
Connectionist systems by programming computers, and, in
particular, in today's Generative AI systems. And, albeit by
implication, I am saying that it doesn't matter how many "...
thousands of scholarly papers" have been "published about
these networks and using that term [neural networks] for
them," makes absolutely no difference. I repeat. Nouning
maketh not the nouned. And, of course, no amount to repeating
the nouning changes this reality, not, at least, in the way I,
and lots of other people, understand the way real things work.
If this makes all my words, and the meanings of them, like
Humpty Dumpty, then I'm happy to be sitting on the wall too,
and happy to fall off it now and again, as we work to further
our knowledge and understanding of real neural systems.
To my claim that we do understand how today's Generative AI
systems are built and work, you say ...
"We know how they work in the sense that we understand
the principles we use to make them, such as back
propagation and the computational solving of partial
differential equations. We understand them at that
level.
"But at another level we scarcely understand them at all.
Hence there is an entire field of research on the
'inscrutability problem' in AI, which I alluded to when I
mentioned that we don't know where or how a Large
Language Model stores its knowledge that Paris is the
capital of France."
Sorry, but we know how these systems are designed and
specified using linear matrix arithmetic, and we know how to
implement this kind of stuff in programmed computers. These
are not new techniques. Matrix arithmetic and it's effective
and efficient implementation is used in plenty of other kinds
of computational work, such as for structural analysis using
finite element techniques -- something I used to do a lot of.
This is all we need to know and understand to build these
systems, and all we need to know to understand how they work
and what they do.
So, to your claims that "... at another level we scarcely
understand them at all," I would ask you to provide some
evidence for this apparently near complete ignorance of how
machines we design and build actually work. What is it you
would say we are not understanding here? I grant you that an
explanation of how ChatGPT, say, generates some particular
text-token output, given some [other] particular input
sequence of text-tokens, would take plenty of doing, but this
does not make it impossible. It just means you'd need plenty
of patience. You can find some quite good visualisations of
this on the web these days.
If, however, as I suspect, your claim is to do with your
assertion that "... we don't know where or how a Large
Language Model stores its knowledge that Paris is the capital
of France" then my response is that you are mistaken, and the
answer is it doesn't store this "knowledge" anywhere.
First, systems like ChatGPT are not simply "so called Large
Language Models". There's loads more machinery inside them
need to make them do what you see them do, most of which is
secret, except in open source cases like DeepSeek -- so go and
take a look! Second, more importantly, these Generative AI
systems do not know, understand, nor reason about, any of what
the stuff they generate -- text, images, video, sound, music,
combinations there of -- appear to you to be about. ChatGPT,
for example, outputs text by generating sequences of
text-tokens. That's all. It does not generate words! That
this text, when you read it appears to you to be about
something, is something you create, not something ChatGPT puts
in the text. This is the illusion all this Generative AI
stuff is built upon. It's Artificial Flower AI, not Real
Light AI: it looks like [human] intelligence, but is not real
intelligence.
So, called artificial neural networks are a class of Machine
Learning techniques. All these systems are in their different
ways, function approximators. These are good for finding
surface features, patterns, and statistical associations, in
the data they are programmed with, so called "trained" on.
These techniques can be, and are, used to build some useful
devices and tools. But they do not construct any kind of
deeper understanding of any underlying causal structures and
mechanisms that exist in what generated the data used to
program them with. There is no magic here, no matter what the
builders of these system may wish for, like to believe, and
sometimes try to tell us they do. An LLM is, at best, a
statistical model of the sequences of text-tokens present in
very large amounts of text that resulted from human writing.
Languaging is a whole lot more than just text-token sequence
generation, which we, the real language using beings here, do
not do, not even when we write.
To end, all the effort on so called 'Explainable AI' is, in my
view, simply a continuation of the misunderstanding of what
Generative AI systems really are, and of what they really do,
and how. Or, they are just further attempts to sustain the
illusion that these systems are intelligent. Let me repeat,
nouning maketh not the nounded. Saying Generative AI systems
know, and understand, and reason, does not mean they do this.
You must show that they do this by showing us how they do
this, and this, I think, from working in AI research, requires
you to first tell us what you mean by knowing, understanding,
and reasoning. And I don't just mean by giving us top level
hand waving descriptions of these notions. I mean precise,
specific, operational definitions, like we do in symbol
processing AI when we build knowledge based systems. Even
Humpty Dumpty knows this!
-- Tim
> On 10 Jan 2025, at 10:28, Humanist <humanist@dhhumanist.org> wrote:
>
>
> Humanist Discussion Group, Vol. 38, No. 312.
> Department of Digital Humanities, University of Cologne
> Hosted by DH-Cologne
> www.dhhumanist.org
> Submit to: humanist@dhhumanist.org
>
<snip>
>
> [2] From: Gabriel Egan <mail@gabrielegan.com>
> Subject: Re: [Humanist] 38.305: AI, poetry and readers (156)
<snip>
>
> --[2]------------------------------------------------------------------------
> Date: 2025-01-08 10:33:51+00:00
> From: Gabriel Egan <mail@gabrielegan.com>
> Subject: Re: [Humanist] 38.305: AI, poetry and readers
>
> Tim Smithers writes:
>
>> . . . we most certainly do know AI machines
>> do not work like human brains do, despite
>> remaining unknowns, perhaps more unknowns
>> than we currently suppose, about how brains
>> are built and function. Why? Because both
>> do not use "neural networks."
>
> Tim goes on to explain this last remark by
> saying that the things in our brains really
> are neural networks but the things in our
> computers are not.
>
> There are two obvious objections to this
> reasoning.
>
> The first is that things don't have to be
> built the same to work the same. Aeroplanes
> work like birds in how they fly: they
> generate lift by deflecting a moving
> airflow over specially shaped wings. Two
> things don't have to be physically
> identical to be functionally similar
> (that is "like" each other, in my
> phrasing) The lenses in my spectacles
> work like the lenses in my eyes, for
> instance.
>
> The second objection is that to say
> that what computer scientists have built
> are not neural networks because they are
> not like brains is begging the question.
> (The question being begged is "what is
> a neural network?")
>
> The analogue electrical device called the
> perceptron was invented to mimic the
> function of the biological device called
> the neuron, and people who now connect
> together layers of perceptrons -- or more
> commonly digital simulations of perceptrons
> -- call the things they make 'neural
> networks'. There are thousands of scholarly
> papers published about these networks and
> using that term for them, so to object
> that they are not really neural networks
> is to risk sounding like Humpty Dumpty
> regarding the meaning of words.
>
> Tim goes on to say that:
>
>> We do know and understand how today's so
>> called Generative AI are built and work.
>> We wouldn't be able to build and operate
>> them if we didn't.
>
> We know how they work in the sense that
> we understand the principles we use
> to make them, such as back propagation
> and the computational solving of partial
> differential equations. We understand
> them at that level.
>
> But at another level we scarcely
> understand them at all. Hence there
> is an entire field of research on
> the 'inscrutability problem' in AI,
> which I alluded to when I mentioned
> that we don't know where or how a
> Large Language Model stores its
> knowledge that Paris is the capital
> of France.
>
> In systems built by the principles
> of Good Old Fashioned AI (GOFAI),
> such as the Expert Systems of the
> 1970s and 1980s, you certainly could
> point to the part of the system that
> contained each bit of knowledge that
> the system held. But a computational
> neural network acquires knowledge
> not by having it explicitly put in
> by a human creator but by ingesting
> a large amount of text and using it
> to tweak a large number of weighted
> connexions between perceptrons,
> and in this process we never
> see where it stores each bit of
> knowledge.
>
> If Tim were right that "We do know
> and understand how today's so
> called Generative AI . . . work[s]"
> (as he writes) then the field of
> research into the inscrutability
> problem and the drive to produce
> 'Explainable AI' would not exist.
> That they do exist argues against
> Tim's position.
>
> Tackling the topic from a different
> angle, Tim argues that human text
> generation involves iterative
> processes of writing and reading:
>
>> Writing is a working out of
>> what to think, and how to think
>> and understand, things we are
>> working on. It's not just a
>> text generation procedure.
>> Writing is a conversation --
>> literally literal -- between
>> us and what the words we read
>> from our own text say to us
>> when we read them, and re-read
>> them, and change them, and start
>> again with them, and thereby
>> discover what we are saying,
>> not say, can say, can't say,
>> and more.
>
> I think anyone who writes professionally
> will agree with Tim's account of the
> iterative process by which humans
> revise their text output to perfect
> it, which machines do not do. But it is
> possible that this iterative process
> is no more than a result of the human
> brain's limitations.
>
> It would seem to be more efficient
> if I could put the 'reading' bit of
> my brain onto the task of checking
> what is being created by the 'text
> generating' bit, all inside my
> head and without having to
> externalize the generated text as
> typed characters and words. But for
> all we know the route out of my brain
> through my arms and hands into pixels
> on a screen and then back in through
> my eyeballs is the only possible
> route because my brain has not
> provided an internal route between
> the requisite parts of itself. The
> fact that minds do text generation
> this way does not indicate some
> special property that machines lack
> and that makes machines inferior.
> The human way may indeed be
> suboptimal.
>
> Regards
>
> Gabriel Egan
>
--[2]------------------------------------------------------------------------
Date: 2025-02-19 17:50:10+00:00
From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 38.316: AI, poetry and readers
Rewinding the tape 2
Dear Jim,
With PhD teaching done for a bit, I thought to go back to your
last post here on "AI, poetry and readers," and respond to
this. To get back in to the conversation I wound the tape all
the way back to where you and me started, and I read all you
and I have posted here on this. We've done quite a lot of
exploring, more than I'd remembered, and sorted through quite
a lot of thoughts and understandings, which I've enjoyed
doing, a lot.
I think it'd be fair to say, we've had a good conversation.
But, though there are more things I could say in reply to you,
I now think your last post leaves our conversation is a good
state, at least for the time being. So, unless there is
something you'd particularly like me to respond to, I propose
we let your last post mark a resting place for our exchanges.
Before doing this, however, I do want to say a big Thank You
to you for your engagement, clear and pressing thinking, and
always friendly manner. Our talking, real languaging, I'll
call it, has certainly been useful to me, and, I hope, to
others here too.
Hurrengora arte. [Until the next time.]
-- Tim
> On 13 Jan 2025, at 06:50, Humanist <humanist@dhhumanist.org> wrote:
>
>
> Humanist Discussion Group, Vol. 38, No. 316.
> 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] 38.313: AI, poetry and readers (58)
>
<snip>
> --[1]------------------------------------------------------------------------
> Date: 2025-01-13 01:28:05+00:00
> From: James Rovira <jamesrovira@gmail.com>
> Subject: Re: [Humanist] 38.313: AI, poetry and readers
>
> Tim -
>
> Thanks for your response, again. I think you're asking good questions about
> what makes for a useful model. So the same thing can be a model for some
> purposes and isn't a model for other purposes.
>
> My one point of disagreement is with what you said about nonsense
> sentences. I think the reaction you describe is for a sentence that isn't
> nonsense. It's only partly nonsense, or mostly nonsense, like Westley was
> "mostly dead" in the *Princess Bride*. Lewis Carroll loves these, but we
> still kind of make sense of them. You can't have any reaction to a truly
> nonsense sentence because it literally makes no sense. Here's the most
> famous one:
>
> Colorless green ideas sleep furiously.
>
> None of these words go together in any coherent way, so it's a nonsense
> sentence. There's no emotional reaction possible.
>
> From here, though, I agree with a lot that you said. I would agree with the
> idea a human mind needs to be present at the reception end, at least, to
> make "meaning." I have always agreed with you that computers aren't
> thinking, and the text they generate doesn't have meaning *to them*. My
> argument is that they can still have meaning *for us as readers *because
> meaning and intentionality is embedded in language itself whether the text
> generator possesses it or not. On a very brute force, material level, words
> on a screen are still and always words wherever they come from.
>
> If I were to think through our examples of sonnets as models, however, from
> the standpoint of a creative writing instructor (I've taught grad and
> undergrad creative writing courses and supervised creative theses, mostly
> in poetry), I could take any of the computer generated sonnets posted here
> and use them as a model for a sonnet because they meet requirements:
> fourteen lines, one of the two most common rhyme schemes (for Elizabethan
> sonnets in the case of those posted here), iambic pentameter, and meeting
> all of those requirements in sensible grammatical units. If you read them
> out loud, they would sound like generally readable and grammatically
> correct sentences. They are technically proficient, in other words.
>
> So if I wanted to show students what a technically proficient sonnet would
> look like, I would indeed use those AI generated sonnets. They serve as
> models for technically proficient sonnets.
>
> HOWEVER, if I wanted to show students something more advanced, subtle, and
> complex, I'd need Shakespearean sonnets, or Petrarchan sonnets, say by
> Petrarch or Wyatt. Those rely more on polysemy and irony to create
> subtleties of meaning that the AI generated sonnets lack. Generally, with
> AI, we never ask, "Did he mean to say this or that?" Or even Billy
> Collins's sonnet that responds to the entire sonnet tradition from the
> point of view of a woman's voice who wishes the poet would quit writing and
> just get her to bed already. AI doesn't think in terms of knowledgeable
> readers who get the joke. My experience has been AI is more likely to
> explain the joke than tell it.
>
> Anyway, thank you for this discussion. Whether we agree or not, and I think
> we agree on quite a bit, you've helped me think through my own ideas.
>
> Jim R
>
<snip>
_______________________________________________
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
