Humanist Discussion Group, Vol. 39, No. 27.
Department of Digital Humanities, University of Cologne
Hosted by DH-Cologne
www.dhhumanist.org
Submit to: humanist@dhhumanist.org
[1] From: Gabriel Egan <mail@gabrielegan.com>
Subject: Re: [Humanist] 39.21: repetition vs intelligence (155)
[2] From: maurizio lana <maurizio.lana@uniupo.it>
Subject: Re: [Humanist] 39.21: repetition vs intelligence (333)
--[1]------------------------------------------------------------------------
Date: 2025-05-21 23:26:39+00:00
From: Gabriel Egan <mail@gabrielegan.com>
Subject: Re: [Humanist] 39.21: repetition vs intelligence
Dear Humanists
I believe that an important bit
of Tim Smithers's account of Word
Embedding algorithms such as
word2vec, as used in Large Language
Models, is mistaken. He rightly
describes the placing of each
"text token" (part of a word)
in its own position within a
high-dimensional space, so that
each has a position described
by a list of coordinates, which
list is 'a vector' in the sense
of that word used in Computer
Science but not in the sense
used in mathematics.
I draw this distinction because
I think the two senses of 'vector'
cause confusion. For Computer
Scientists any list of numbers
can be called a vector. But
in mathematics, a vector is
a quantity that has magnitude
(size) and direction but not
position. This difference
matters. A Computer Scientist
speaking loosely will use the
term 'vector' for the coordinates
(1,1,1), which identify the
corner of a unit cube that is
furthest from the origin (at
0,0,0).
But in mathematics, where vectors
do not have positions, the
numbers (1,1,1) are a vector in
the sense of being a displacement
from wherever you are now, say
at point (2,5,3, to a new point,
here (3,6,4). For mathematicians
(1,1,1) as a vector means 'move
one unit along the x axis, one
unit along the y axis, and one
unit along the z axis', with
no reference to where you started.
Smithers blurs this distinction when
he writes that in word-embedding
algoritms:
<<
Text tokens which frequently occur
close to each other in sequences of
text tokens in the text used to
program these systems -- so called
"train" them -- have places in this
vector space which are close together.
>>
That is not right. Words end up
close to one another in vector
space not because they are "close
to each other" in the training
text. Rather, they are close
to one another in the vector
space if the training process
identifies that they are close
in meaning.
But here even the notion of
'close' is misleading, because what
really matters in the high-dimensional
vector space used in LLMs is the
mathematical notion of a vector
as a displacement from one place
to another. These displacements
really do encode meaning, because
the dimensions themselves encode
meaning.
What matters is not that 'king' and
'queen' are "close to each other"
in the training text (they probably
aren't). What matters is that
they get used in similar contexts:
the things we say of kings are
like the things we say of queens,
but with the gender flipped.
And what is so remarkable about
the dimensions of the high-
dimensional vector space of LLMs
is that there will be a dimension
(or more than one) that encodes
something like what we mean by
gender. Thus there will be a
vector in the mathematical sense,
that is, a particular displacement
along particular dimensions and
not others, that gets you from 'king'
to 'queen'.
We know that this vector/displacement
encodes gender because if we add
the same vector/diplacement to
'trousers' we end up close to the
points for words such as 'dress' and
'skirt' and if we add the same
displacement to 'penis' we end up
close to the point for the word
'vagina'.
That is, the displacement (the
mathematical sense of 'vector')
really does encode something
much like our sense of 'gender',
and it is transportable to
any part of the high-dimensional
space. Of course, from some
starting points this displacement
takes you nowhere interesting:
add the 'male-to-female'
displacement to 'lemonade' and
you don't land anywhere
interesting, because gender
doesn't apply to lemonade.
But add it to 'emeritus' and
you do land near to 'emerita'.
I think that it is confusion about
what vectors are in the vector space
that LLMS use that leads Smithers
to write that:
<<
No semantics of any kind plays
any role in the computations
carried out by the LLM on these
numerical vectors.
>>
Clearly, if a certain vector (in
the mathematical sense of a
displacement) is a general-purpose
male-to-female gender flipper,
then that vector is encoding
something semantic. How far apart
words are in the training text
really isn't what this is about.
Regards
Gabriel Egan
--[2]------------------------------------------------------------------------
Date: 2025-05-21 19:47:04+00:00
From: maurizio lana <maurizio.lana@uniupo.it>
Subject: Re: [Humanist] 39.21: repetition vs intelligence
thank you for these lines Tim.
their the peak is here, for me:
> Text is the marks left by some human writing, and, now-a-days,
> often printed or screen rendered using suitable well designed
> font(s) and typographical designs. Text is not the same as
> words. The words involved were formed in the head of the
> author and remain there. Writing words to say something
> involves encoding the chosen words in some shared alphabet and
> shared spelling and grammar. This results in the marks we
> call text. Text is thus a sequence of signs, and it must be
> read, by, of course, something that can read these signs, to
> re-form the words of the author. These again formed words are
> formed in the reader's head, they are not found and somehow
> picked out of the text; the signs are not the words, they are
> signs for words. This notion of "picking up the words" is not
> what reading is, though this is how it might seem to us, and
> how we often talk about it being. This confusion -- the text
> is the words -- was harmless when we [just about] only had
> text from human writing, but now we have, thanks to things
> like ChatGPT, automated text generation systems, and lots of
> text which is not the result of any kind of writing. Just
> because we can read this automatically generated text, and
> form words in our heads from this reading, words which mean
> something to us, and thus give us the impression that the text
> is about something, does not mean, nor necessarily make, the
> generator of this text a writer. To be a writer requires the
> author to be a reader of the written text, and, or course,
> lots of other text. And it requires the writer to have a mind
> in which they form words to say something with. ChatGPT, and
> other Generative AI systems like it, do not read anything.
> ChatGPT does no reading of your [so called] prompt. The text
> you make by writing your prompt is simply chopped into a
> sequence of text tokens which are, in turn, used to build a
> sequence of vector encodings, together with quite a lot of
> other stuff added to your prompt text by the always hidden
> prompt processing ChatGPT has to do. (ChatGPT is not just an
> LLM, it has plenty of other machinery needed to make it do
> what it does.)
and just this evening saw this article:
Porter, Brian, e Edouard Machery. «AI-generated poetry is
indistinguishable from human-written poetry and is rated more
favorably». /Scientific Reports/ 14, fasc. 1 (14 novembre 2024): 26133.
https://doi.org/10.1038/s41598-024-76900-1.
allow me to say that literature is like cuisine: if you are not educated
you are not able to distinguish the flavors, and to appreciate them and
their combination; or you are educated only in some flavors (e.g. the
Italian people who abroad first of all search for an Italian restaurant).
if you are not educated in literature you are not able to distinguish
and fully appreciate it. you could end tasting and appreciating junk
food without even knowing what junk food (junk information) is.
so my question, as a professor and as a citizen is: what do we do/what
can we do/what should be do to help people to understand and appreciate
good food for the mind?
Maurizio
Il 20/05/25 17:14, Humanist ha scritto:
> Humanist Discussion Group, Vol. 39, No. 21.
> Department of Digital Humanities, University of Cologne
> Hosted by DH-Cologne
> www.dhhumanist.org
> Submit to:humanist@dhhumanist.org
>
>
>
>
> Date: 2025-05-16 09:19:21+00:00
> From: Tim Smithers<tim.smithers@cantab.net>
> Subject: Re: [Humanist] 39.4: repetition vs intelligence?
>
> Dear Willard,
>
> I'm behind Jim Rovira's "we're not talking about the same
> thing when we talk of intelligence here," and I like
> Maurizio's comments on repetitions, and I greatly enjoyed
> Manfred's explorations of what gets mixed up and confused in
> our conversations on and around Generative AI. And, I'd like
> to add to these, if I may, Manfred?
>
> Our discussions about Generative AI systems, so called Large
> Language Models (LLMs), and what these systems do and don't
> do, to be useful, needs, I think, more precision and
> discipline. Else our conversations are empty of needed
> clarifications and right understandings, and are like a group
> of people batting a ballon around. Entertaining for a while,
> perhaps, but of no useful consequence.
>
> The word 'intelligence' is what I call an ice-hockey-puck
> word. It can easily be pushed and knocked around the semantic
> playing surface to mean quite different things, and thus used
> to score "goals" in all sorts of conversations. This is the
> utility of words like this, but we can never pin anything down
> with these ice-hockey-puck words. Languaging with
> ice-hockey-puck words results in what I call Ice Rink talk:
> conversation happily slides all over the place going nowhere.
> [I used to call these shove-ha'penny words, and the use of
> such words ha'penny talk, but people stopped understanding
> what I was talking about. Who here played shove ha'penny as a
> kid?]
>
> To get somewhere useful -- a place of greater clarity, sounder
> understanding, and, perhaps, better identified disagreement --
> any use of the word 'intelligence' must come with some
> elaboration of what is to be meant by this word by the author
> who choses to use it, and thus, what is to be understood from
> it by readers: what, for the purposes of the conversation are
> we to understand 'intelligence' to consist in? This could be,
> for example, the intelligence of making winning chess moves,
> or the intelligence of writing a gripping and fear inducing
> ghost story, or the intelligence of building a coherent
> historical account of something which happened, but for which
> we don't have complete and detailed records, or it could be
> the intelligence of bike riding in busy urban settings, and
> mending the punctures when they happen. Of course, there are
> endless more examples like this ... so you'll happily excuse
> me for not listing them, I trust. Intelligent of me, no?
>
> 'Creative' is another ice-hockey-puck word. [Ha'penny word
> sounds so much better, no?] Using 'intelligence' and
> 'creative' together results in what I call a Snooker
> conversation; these two words can be bounced off each other is
> all sorts of ways, and go off in all sorts of directions on
> the semantic playing surface. Fun, perhaps. Useful? Hardly.
>
> In today's super-hyp'ed talk about Generative AI systems other
> words, which 'til now had good, strong, mostly commonly
> understood, meanings, have had their semantic tethers cut.
> Words such as 'knowing,' 'reasoning,' 'understanding,'
> 'writing,' 'hallucination,' and more. We can now, it seems,
> say whatever thing we like does knowing, reasoning,
> understanding, and writing, and demand this thing really
> really does know, reason, and understand, and write, just like
> humans know, reason, understand, and write, only better, and
> nobody can have good reason to contradict these obviously
> false assertions. This is speech acts gone mad. Yet we seem
> to be happy to accept this madness. Why?
>
> I'm currently teaching a PhD course on making and using models
> in research, to PhDers from across all the disciplines in the
> Arts, Egineerings, Humanities, and Sciences. So, the term
> 'Large Language Model' (LLM) gets talked about some.
>
> You're correct, Willard, to preface this term with "so
> called." LLMs are only taken to be models of language because
> their builders call them models of language. Which, in a less
> mad world, does not make them models: naming maketh not the
> named. They are not models of language because none of these
> people have shown [the rest of us] how their LLM constructions
> actually model language in some useful way. LLMs do not deal
> in words, not really. All the simple minded explanations of
> how LLMs "predict" the next word in some sequence of words are
> Noddy explanations which seriously mislead people into
> thinking LLMs deal in and process words. They don't. They
> process text tokens, and the large majority of the text tokens
> these systems use are not words; they are bits of text. See,
> for example, "ChatGPT’s vocabulary" here
> <https://emaggiori.com/chatgpt-vocabulary/>, but read enough
> to finally get to where what ChatGPT really uses is explained;
> it starts off with the usual misleading story about word
> predicting.
>
> ChatGPT uses more than 100,000 different text tokens, and
> these are encoded using UTF-8 and represented using very large
> numerical vectors to form what is called the "embedding
> space." This is often [grossly mis-] described as the
> "semantic space" of the LLM. No semantics of any kind plays
> any role in the computations carried out by the LLM on these
> numerical vectors. It's just made to look like it does, if
> you play fast and lose with what is meant by semantics and how
> words mean anything. Text tokens which frequently occur close
> to each other in sequences of text tokens in the text used to
> program these systems -- so called "train" them -- have places
> in this victor space which are close together. If you present
> this using text tokens we recognise, and thus [automatically]
> read as words, it can look like this "closeness" in the text
> token vector space "captures" the semantic relationships of
> the words involved. But this semantic relationship is an
> artefact of the statistics of text token patterns found in the
> text used to program the system. To claim this token vector
> space is a semantic space of words would require showing that
> this is the only kind of relationship found between the
> represented tokens, not just something we can find if we pick
> the right text tokens to use, which, of course, cannot be true
> since most of the text tokens represents are not words. LLMs
> do capture the statistics of text token relationships found in
> the original text, but given how they are programmed with all
> this text, they cannot do anything else. Claiming this
> statistics of text token patterns is the same as the semantics
> of word sequences is either deliberate deception or ignorance
> induced delusion on the part of the people to claim this.
>
> So, the best we might be able to say about LLMs is that they
> are statistical models of text token patterns found in a
> ginormous collection of texts that have resulted from some
> human writing, after ripping all images and other non-text
> content that are integral components of the original text, and
> ripping out all the typographical formatting of this original
> text needed to make it readable by us: try reading long
> sequences of UTF-8 codes!
>
> Text is the marks left by some human writing, and, now-a-days,
> often printed or screen rendered using suitable well designed
> font(s) and typographical designs. Text is not the same as
> words. The words involved were formed in the head of the
> author and remain there. Writing words to say something
> involves encoding the chosen words in some shared alphabet and
> shared spelling and grammar. This results in the marks we
> call text. Text is thus a sequence of signs, and it must be
> read, by, of course, something that can read these signs, to
> re-form the words of the author. These again formed words are
> formed in the reader's head, they are not found and somehow
> picked out of the text; the signs are not the words, they are
> signs for words. This notion of "picking up the words" is not
> what reading is, though this is how it might seem to us, and
> how we often talk about it being. This confusion -- the text
> is the words -- was harmless when we [just about] only had
> text from human writing, but now we have, thanks to things
> like ChatGPT, automated text generation systems, and lots of
> text which is not the result of any kind of writing. Just
> because we can read this automatically generated text, and
> form words in our heads from this reading, words which mean
> something to us, and thus give us the impression that the text
> is about something, does not mean, nor necessarily make, the
> generator of this text a writer. To be a writer requires the
> author to be a reader of the written text, and, or course,
> lots of other text. And it requires the writer to have a mind
> in which they form words to say something with. ChatGPT, and
> other Generative AI systems like it, do not read anything.
> ChatGPT does no reading of your [so called] prompt. The text
> you make by writing your prompt is simply chopped into a
> sequence of text tokens which are, in turn, used to build a
> sequence of vector encodings, together with quite a lot of
> other stuff added to your prompt text by the always hidden
> prompt processing ChatGPT has to do. (ChatGPT is not just an
> LLM, it has plenty of other machinery needed to make it do
> what it does.)
>
> So, to mend the usual, and still needed, semantic tether the
> word 'writing' used to have, ChatGPT does not, and cannot,
> write, it only generates text. It has no mind in which to
> form words and then work out how to writes down using signs we
> can read. It does not, and cannot, read. It has no mind in
> which to form words, it chops text into text tokens, a
> different system of signs, and not one we can read.
>
> To take the text generated by systems like ChatGPT as writing,
> and to take this writing to be the result of something that
> works out something to say, and then works out which words to
> use to say this with, and then writes these words for us to
> read, is to hallucinate.
>
> To hallucinate : to experience an apparent sensory
> perception of something that is not actually present.
>
> This is the real meaning of to hallucinate, a meaning we
> clearly still need it to have. Generative AI system do not
> hallucinate. And, the only fabrication, the only "making
> things up," they do is the fabrication of sequences of text
> tokens. They do not say anything by writing, but they are
> built to make it look like they do. This is deliberate
> deception, a kind of dishonesty.
>
> Only humans write, machines only generate text. Let's try to
> keep this simple and evident distinction in the way we talk
> about these things.
>
> -- Tim
>
>
>
>> On 8 May 2025, at 09:38, Humanist<humanist@dhhumanist.org> wrote:
>>
>>
>> Humanist Discussion Group, Vol. 39, No. 4.
>> Department of Digital Humanities, University of Cologne
>> Hosted by DH-Cologne
>> www.dhhumanist.org
>> Submit to:humanist@dhhumanist.org
>>
>>
>>
>>
>> Date: 2025-05-08 06:51:41+00:00
>> From: Willard McCarty<willard.mccarty@mccarty.org.uk>
>> Subject: repetition vs intelligence
>>
>> This is about the current state and probable trajectory of artificial
>> intelligence, I'd hope without promotional futurism. (Prominence of the
>> future tense in writings about AI I find very interesting indeed, but
>> here it is, at least for me, only a distraction.)
>>
>> My question is this: to what extent, in what ways, do the strategies of
>> the so-called Large Language Models produce results that only echo back
>> to us current linguistic behaviour (parole), in effect saying nothing
>> new, however useful, however news to the questioner? The current term
>> for the misbehaviour of LLMs when they make things up seems to be
>> 'hallucination'; far more accurate would be 'fabrication'.
>> Hallucinations are much more interesting, but used of LLMs lets them off
>> the hook.
>>
>> We could say, as a friend of mine did, that saying something new in my
>> sense, i.e. being truly creative, is exceedingly rare. But isn't that
>> exactly what we want of intelligence? What would the artificial kind
>> have to do to qualify? Or do we have examples, are they being noticed
>> and investigated?
>>
>> Enough for now, I trust. Comments eagerly welcomed!
>>
>> Best,
>> WM
>> --
>> Willard McCarty,
>> Professor emeritus, King's College London;
>> Editor, Humanist
>> www.mccarty.org.uk
non ci sono altri giorni
che questi nostri giorni
italo calvino, il cavaliere inesistente
------------------------------------------------------------------------
Maurizio Lana
Università del Piemonte Orientale
Dipartimento di Studi Umanistici
Piazza Roma 36 - 13100 Vercelli
_______________________________________________
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
