Chris von Csefalvay is a computational epidemiologist/data scientist working at the intersection of AI, epidemiology and public health.

Chris von Csefalvay

The year is 1959. Eisenhower is on his second term, Castro just kicked Batista out of the country and Ray Charles’s Let the Good Times Roll is topping the charts.

The 95% myth

Over at the work blog, I’m discussing what knowledge means for large language models (LLMs), and the ways in which we can leverage this knowledge dividend for better inference.  Read the full post here.

The knowledge dividend of large language models

The awesome thing about language is that, well, we all mostly speak it, to some extent or another. This gives us an immensely powerful tool to manipulate transformational tasks. For the purposes of this post, I consider a transformational task to be essentially anything that takes an input and is largel intended to return some version of the same thing. This is not a very precise definition, but it will have to do for now.

LAIR - Language As Intermediate Representation

As the new Omicron variant of concern of COVID-19 emerged over Thanksgiving week (of all times!), attention has been turning to Paxlovid (PF-07321332). If, like me, you torment yourself by perusing conspiracy theories, you might have heard it referred to as Pfizermectin.

Peace, love, Paxlovid and ‘Pfizermectin’

Tip  If you’re here for the DOI matching script, the link is here. Quarto is a great tool for reproducible research. It is also a great tool for building websites. After about a decade of Wordpress-based websites, I’ve moved to Quarto primarily for two reasons. First, most of my content is largely static.

Quarto project scripts are awesomeness

There’s a notion in artificial intelligence known as Moravec’s paradox: it’s relatively easy to teach a computer to play chess or checkers at a pretty decent level, but near impossible to teach it something as trivial as bipedal motion. (Hassabis 2017) The sensorimotor tasks that our truly wonderful brains have mastered by our second birthday are much harder to teach a computer than something arguably as ‘complex’ as beating a chess grandmaster.

The hardest AI problem you’ve never heard of.

It appears that in what is clearly a wonderful little PR stunt, a Polish rum company managed to do a Sophia and appoint an ‘AI-driven’ ‘robot’ as its ‘CEO’.  The other guilty party to this pile of steaming bovine excrement is Hanson Robotics, famous for giving us Sophia, the “world’s first robot citizen”. Most of what I’m saying here goes just as well for Sophia.

The skeuomorphic fallacy

Yesterday, I completed a marathon on the SkiErg, with a pretty respectable time (see results here). I know I can do better, and I know I

 will

do better. On the other hand, this was a marathon where I pretty much did everything wrong. And that was sort of an instructive experience in its own right. The title is somewhat misleading. Not much went wrong – the marathon itself was smooth sailing.

How to do a SkiErg marathon entirely the wrong way (but still finish)

There’s something special about language. It is ‘our own’, it is ‘us’, in a profound way, and quite surprisingly, more so than art. I was deeply struck by this when I first saw reactions to large generative language models that created realistic, human-ish prose.

Beyond Broca

There might be valid concerns about the mRNA vaccines. Then there are the clearly insane ones, such as the claims that it involves the injection of a microchip. The argument that mRNA vaccines are a form of ‘gene therapy’ are somewhere in the middle, and that makes responding to it somewhat difficult.

Why the COVID-19 vaccines are not gene therapy

The temple of Asclepios, the Greek god of healing arts and medicine, at Epidaurus was pretty much the ancient Greek world’s equivalent of the Mayo Clinic. It then tells us a lot about the Greek worldview of healing that one of the things the temple complex prominently featured was a theatre. The Greeks believed in drama therapy, but in perhaps a slightly different way of how we think of it today.

The Moral Pulse of the Machine

Every year, the Commandant of the Marine Corps publishes a reading list of books that often only bear on warfighting tangentially at best. The idea behind this is that those entrusted with the lives of servicemembers should have an understanding of the world that goes beyond the profession of arms. In much the same way, I have been advising data scientists to go beyond professional literature.

Five non-data-science books for data scientists

I used to be on Quora. I’m not anymore. But I used to be. And I wrote some answers. Here are some of my favourites, categorised by field… sort of.

 AI

If I wrote an AI and trained it to score 200 on a standardized IQ test, what could a human of 100 IQ still do better? Read the full answer here. Will true AI come before the ability to completely simulate a human brain? Is there a difference? Read the full answer here.

My favourite Quora answers

In my latest post for the Starschema blog, I discuss the end of dashboards, and what comes next:  Read the full post here.

As we may see: the world after dashboards

In the middle of the green and pleasant countryside of England lies a small piece of commons –– land, that is, over which a number of people share certain rights. In this hypothetical scenario, it’s six farmers,  to . Each of them has a hypothetical cow, which differs from ordinary cows in that it grazes a fixed unit, and produces a fixed amount of milk (unlike real cows, which vary wildly in their input and output parameters, or so I’m told).

Ransomware attacks as n-person prisoner’s dilemmas

Say you’re busing tables and you’re trying to pass someone in a wheelchair. What do you do? Do you say “excuse me” and wait for them to move? Do you say “excuse me” and then try to pass them? Do you just try to pass them? Do you say nothing and just try to pass them? All of these are, actually, pretty legitimate answers. Now, say you’re a robot.

What I learned from getting bodied by a robot.

Asemantic Induction of Hallucinations in LLMs

There’s a style of visual design I’m inordinately fond of called Raygun Gothic. It’s hard to describe what the hell exactly one needs to be on to enjoy it, but think of it like the aesthetic from the latter Fallout games with a more optimistic outlook on the future. Gibson described it as “the future that never was”, and I think that’s a pretty apt description. We get these competing futures in technology all the time.

Prompt Engineering: The Art of Yesterday

It’s not every day that you find out you have climbed the exalted heights of another discipline. My work is pretty interdisciplinary, but it shocked me, too, that I’m apparently holding forth on neoliberalism and the epistemic question in African universities (archive link):  This, of course, came at some surprise to me, as I have never written anything on the topic.

Stochastic parrots, cap and gown edition

In the first four entries (1 2 3 4) of this sequence, I have focused primarily on what LLMs aren’t, can’t, won’t, wouldn’t and shouldn’t. It’s probably time to conclude this series by that much awaited moment in all stories, where the darkest night finally turns into a glorious dawn, where we finally arrive at the promised land, where we finally get to talk about what LLMs

 could

be. What I see as the most successful potential model of

Teams of Rivals

My recent post on the Starschema blog discusses the need for better data products to tackle future pandemic challenges:  Read the full post here.

Data for the next pandemic

I have probably spent more time looking at Poussin’s

 Dance to the Music of Time

than any other work of art. Sneaking off to the Wallace Collection in London and just looking at the

 Dance

was my comfort activity while living in London – a time that was not exactly devoid of its trials. It’s not, by any measure,

 great

art, insofar as such judgments can be made with any objectivity.

The Lyre of Hephaestus

In a recent paper that has attracted the interest of popular media as well, Fabio Urbina and colleagues examined the use (or rather, the abuse) of computational chemistry models of toxicity for generating toxic compounds and potential chemical agent candidates.(Urbina et al. 2022) Urbina and colleagues conclude that Urbina, Fabio, Filippa Lentzos, Cédric Invernizzi, and Sean Ekins. 2022.

A different shade of grey

I love posts that allow me to merge some of my addictions.

Auto-DOI for Quarto posts via Rogue Scholar

I have probably spent more time looking at Poussin's *Dance to the Music
of Time* than any other work of art. Sneaking off to [the Wallace
Collection](https://wallacelive.wallacecollection.org/eMP/eMuseumPlus?service=direct/1/ResultListView/result.t1.collection_list.$TspTitleImageLink.link&sp=10&sp=Scollection&sp=SfieldValue&sp=0&sp=0&sp=2&sp=SdetailList&sp=0&sp=Sdetail&sp=0&sp=F&sp=T&sp=2)
in London and just looking at the *Dance* was my comfort activity while
living in London -- a time that was not exactly devoid of its trials.
It's not, by any measure, *great* art, insofar as such judgments can be
made with any objectivity. Like most French Baroque allegorical
paintings, it's considered somewhat 'boring'. It neither has, nor claims
to arrogate to itself, the fiery passion of a Van Gogh, the
compositional genius of a Vermeer or the turbulence of a Turner. It's
not particularly well known, and indeed more know it for having lent its
name to Anthony Powell's dodecalogy (quite incidentally, another of my
favourite works of art). In as far as there can be a sensible debate as
to what is, and isn't, 'good' or 'great' art, few would consider *Dance*
to fall into that category. What nobody would argue, however, is that it
*is* a piece of art.

::: {.quarto-figure .quarto-figure-center}
<figure class="figure">
<img
src="https://chrisvoncsefalvay.com/posts/lyre-of-hephaestus/IMG_8537.jpeg"
class="img-fluid figure-img" />
<figcaption>Nicolas Poussin, Dance to the Music of
Time.</figcaption>
</figure>
:::

What, then, if I gave a somewhat clumsy description of the scene to
Stable Diffusion, the currently fashionable model of generating images?
Let's try this:

> A painting, in the French Baroque style, of four female figures,
> allegories of the four seasons. They hold each other by the hand,
> dancing in a circle. Time, represented by an old man, plays a lyre.
> The scene is set in the early morning.

Here's what I got:

::: {#fig-poussin-stablediffusion .quarto-figure .quarto-figure-center .anchored}
<figure class="figure">
<img
src="https://chrisvoncsefalvay.com/posts/lyre-of-hephaestus/IMG_8536.jpeg"
class="img-fluid figure-img" />
<figcaption>Figure 1: Stable Diffusion’s attempt at my description of
the Dance.</figcaption>
</figure>
:::

It's not great by any stretch of the word. The faces, like in most
diffusion models, are a mess. None of the figures get the allegory of
time and seasons (of the year, but also of life, and of stations of an
individual's life). Certainly some more/better prompting could have
gotten a slightly better result, but that's really not the point.

The point is that it's not *art* not because it's not 'good enough' to
be art, but because it is fundamentally, qualitatively different.

::: {#one-art-please .section .level1 .page-columns .page-full}
# One art, please

At this point, we need to introduce a key concept to the argument to
follow, namely that of a simulacrum. For the purposes of this argument,
I shall define a simulacrum as follows: something that resembles the
product of an intrinsically human activity but which can be obtained
without that activity, including by non-human actors, such as non-human
animals or, as it may be, artificial intelligences.

::: {.quarto-figure .quarto-figure-center}
<figure class="figure">
<img
src="https://chrisvoncsefalvay.com/posts/lyre-of-hephaestus/IMG_8534.jpeg"
class="img-fluid figure-img" />
<figcaption>One art, please!</figcaption>
</figure>
:::

::: {.page-columns .page-full}
At the heart of the notion of a simulacrum is that the product and the
process are different things. Art is not a product, but a process. I am,
as I will explain later, not singling out art in some romantic
exaltation of the artistic as quintessentially human, nor am I limiting
art to what one might find in a museum. When I discuss art, I fully
intend to include a toddler's finger painting and even some
absent-minded doodles on the margins of a less-than-gripping journal
article.^1^ There are other activities that may become the subject
matter of a simulacrum, language being perhaps of foremost interest (on
which, once again, see *infra*). And just as Zoidberg can't buy 'art',
merely a 'work of art' ('work' here having the meaning of 'product' --
an 'artwork' or 'work of art' being not something with 'art
characteristics' but rather the result of the activity we know as
'art'), an artificial intelligence or a non-human animal cannot create
'art', either.

::: {.no-row-height .column-margin .column-container}
^1^ Because who of us hasn't been bored enough by a dull paper to start
doodling?
:::
:::

At the heart of this is the difference between, of course, the process
and the product. Artificial intelligence generally works according to a
very simple inductive principle: given a bunch of examples (such as
pairs of images and descriptions of images), learn the parameters of a
function ![](https://latex.codecogs.com/png.latex?f(d)%20=%20I) so that
for any arbitrary description
![](https://latex.codecogs.com/png.latex?d), an image
![](https://latex.codecogs.com/png.latex?I) is generated so as to
minimise a loss function
![](https://latex.codecogs.com/png.latex?%5Csum_%7Bi=0%7D%5EN%20J(d_i,%20I_i))
over the ![](https://latex.codecogs.com/png.latex?N)-sized training set.
![](https://latex.codecogs.com/png.latex?f(d)) may be quite complex
indeed, and ![](https://latex.codecogs.com/png.latex?J(d,%20I)) may
comprise various steps up to and including human input (reinforcement
learning with human feedback -- RLFH for short --, which is how most of
the LLMs currently in vogue were trained, include this step quite
prominently), but in the end, it's all about learning
![](https://latex.codecogs.com/png.latex?f). For a neural network, for
instance, this 'learning' takes place by way of an algorithm called
backpropagation, which changes the values of weights and bias matrices
that form part of ![](https://latex.codecogs.com/png.latex?f) so as to
minimise ![](https://latex.codecogs.com/png.latex?J). Regardless of how
![](https://latex.codecogs.com/png.latex?f) is learned, what is at the
heart of the entire process is learning a function that simulates the
outcome. Stable Diffusion, for instance, is not concerned with
simulating the *process* of creating art, it is concerned with
simulating the *outcome* of creating art.
:::

::: {#human-processes .section .level1 .page-columns .page-full}
# Human processes

That matters, because not everything is about outcomes. In fact, as
humans, we recognise that some of the most important things *aren't*.
Consider moral judgment, the individual and societal assignment of a
moral character to human actions. Yes, outcomes do matter, but they
aren't the be-all end-all of moral judgment. We differentiate between
murder, manslaughter, reckless homicide, excusable homicide and
justifiable homicide, even though the end result is the same
(i.e. someone is dead), because we acknowledge that different levels of
moral agency, and different forms of its exercise, are at play in each
of these. We do so because we recognise humans are more than mere
outcome generating machines: how (and why) something came to be is often
just as important, if not more so, than what the end result is.

::: {.page-columns .page-full}
This is a uniquely human differentiation. In the same vein, then, there
are things that only humans are capable of. A machine can certainly
*take a life*, but it cannot *commit murder*, because the latter is
defined not just by an outcome but also the mental state (*mens rea*)
that led up to it. We think of Xerxes flogging the sea as the act of a
megalomaniacal despot who has taken leave of his senses precisely
because we recognise that punishing the sea, incapable of moral
responsibility for its stormy waves washing away Xerxes's pontoon
bridges, as preposterous. Punishment requires, if not guilt, at least a
subject *capable* of being guilty. Punishing the innocent is unjust, but
punishing something incapable of being guilty or innocent or otherwise
morally responsible is worse -- it is *insane*.^2^

::: {.no-row-height .column-margin .column-container}
^2^ Not that otherwise quite rationalistic societies did not engage in
trials of lifeless things. [Walter Woodburn Hyde wrote a wonderful paper
on
this](https://scholarship.law.upenn.edu/cgi/viewcontent.cgi?referer=&httpsredir=1&article=7589&context=penn_law_review).
:::
:::

And just as moral guilt is uniquely human, so is creativity. This is not
to suggest an exalted role for creativity -- much as we tend to think of
that term to have a positive meaning, for the purposes of this argument,
creativity is any act of the mind that brings forth something. This
comprises not just things like creating art, but also using language.
There are many functions of language, ranging from the informative
through the performative (what Searle called 'speech acts') to the
phatic. However, humans communicating always do so with *some* intent.
That's different from machines using words. A machine uses words, and
maybe uses language, but it does not 'speak', for that requires an
intent, something no machine, howsoever sophisticated, can create. Like
art, speech is not just about *what*, but the *why*, of an action. There
are many reasons my neighbour greets me with "hey, how're you doing?" --
maybe it's genuine interest in how I am doing, maybe it is as a kind of
social bond-strengthening (phatic communication), but it is backed by
some sort of intent that goes beyond just wanting to produce words.
:::

::: {#the-tune-without-the-music .section .level1 .page-columns .page-full}
# The tune without the music

A simulacrum, then, is what's left when the uniquely human *process* is
removed, and only the outcome remains, obtained through a different
process (e.g. inference from a model generated by way of minimising the
loss function over a training set). It is the tune without the music. A
simulacrum doesn't necessarily have to be 'worse' than the real deal --
quite likely, Stable Diffusion can produce a better drawing than I
could, and GPT-4 might well write a better poem than my hamfisted
attempts at poetry. However, neither of those *are* the real thing.
Stable Diffusion will be able to do better drawings than I do, but it
won't ever be able to do better *art*, because it'll never be able to do
art.

In that sense, it's the opposite of 'going through the motions': a
simulacrum is an often indistinguishable in its end product from the
'real thing', but arrives there by a different route. This is, indeed,
why it is so deceptively similar to the 'real thing', and what gives
rise to perhaps its darker side. In a recent interview on one of my
favourite podcasts, [Machine Learning Street
Talk](https://www.youtube.com/c/MachineLearningStreetTalk), Daniel
Dennett referred to the simulacrum of human communicational interaction
as "counterfeit people". I disagree with some of Dennett's position, but
this is worth watching in its entirety:

::: {.quarto-video .ratio .ratio-16x9}
::: {.iframe}
::: {#player}
:::

::: {.player-unavailable}
# Er is een fout opgetreden. {#er-is-een-fout-opgetreden. .message}

::: {.submessage}
[Bekijk deze video op
www.youtube.com](https://www.youtube.com/watch?v=axJtywd9Tbo){target="_blank"}
of zet JavaScript aan als dit uitstaat in je browser.
:::
:::
:::
:::

::: {.page-columns .page-full}
I think Dennett is wrong on a few points,^3^ but the language of
'counterfeits' gets some of this point across quite well. A counterfeit
currency isn't really all *that* different from an authentic bank note.
What is different is the intent by which it is created. The counterfeit
lacks the moral intent that it be honoured in return (that's why people
forge currency -- there would be little point in counterfeiting bank
notes if one actually intended to honour them!).^4^

::: {.no-row-height .column-margin .column-container}
^3^ Importantly, I think he's falling victim to the trap he so clearly
identifies. The lack of moral agency of AI, which would make these
pretend-humans something other than 'real' is exactly why the semantics
of 'counterfeiting', which involves deception, an act of quintessential
moral agency, is altogether inappropriate. But that's for another day.

^4^ Arguably, all modern fiat currency is at least a little fake, but
that is *also* another argument for another day.
:::
:::
:::

::: {#the-heart-of-the-matter .section .level1 .page-columns .page-full}
# The heart of the matter

In the end, outcomes are *not* all that make us who we are. The poet,
the artist, the murderer and the thief -- they are who and what they are
not solely because of outcomes, but because of what led up to those
outcomes. And that's ultimately the most human thing any of us will ever
get to experience: to be someone not because of an outcome that a
machine or an algorithm could feasibly mimic, but because of the
quintessentially human process of decisions, good or bad, actions,
passion, foibles. The things that make us who we are are in turn a
*product* of who we are.

::: {.page-columns .page-full}
And that's what gives meaning to those outcomes, meaning that a
machine-generated limerick could never aspire to -- not because it is
not 'good enough', but because it is just a qualitatively different
thing. It is, to use Pauli's turn of phrase, not that the Stable
Diffusion-generated painting in Figure 1 is 'not right' (or
aesthetically pleasing, in this case) -- it's that it's 'not even
wrong'. It lacks the aptitude of being good or bad art, no matter how
good or bad (likely the latter) it would be if it didn't. It could, in
theory, be good enough to put Poussin to shame, and it would still be no
closer to being art. It could be Turner Prize-winning work,^5^ and yet
be put to shame by a toddler's finger-painting.

::: {.no-row-height .column-margin .column-container}
^5^ Let's, for the time being, entertain the notion that that still
means *something*.
:::
:::

A world in which the poet and the machine are on equal footing is a
world that, for all intents and purposes, does not have poets anymore.
It is a world in which outcomes determine everything, and that makes it
a world that is not worth inhabiting: for a world of outcomes alone is a
world in which our choices, our actions, our most irreplicably human
strengths and weaknesses alike play no role.

But that is not the world we inhabit -- thankfully. We -- still? -- live
in a world in which the way we arrive at something, be it a poem or a
painting or a dead body, matters. And as long as that is so, the
simulacra of human action will remain inhabiting a different world of
their own. They may reach a kind of technical verisimilitude, but not a
moral verisimilitude.

And *that* is what makes all the difference.
:::

::: {#quarto-appendix .default}
::: {.section .quarto-appendix-contents}
## Citation {#citation .anchored .quarto-appendix-heading}

<div>

::: {.quarto-appendix-secondary-label}
BibTeX citation:
:::

``` {.sourceCode .code-with-copy .quarto-appendix-bibtex}
@misc{csefalvay2023,
 author = {Chris von Csefalvay},
 title = {The {Lyre} of {Hephaestus}},
 date = {2023-10-16},
 url = {https://chrisvoncsefalvay.com/posts/lyre-of-hephaestus},
 doi = {10.59350/pd0q6-r5n60},
 langid = {en-GB}
}
```

::: {.quarto-appendix-secondary-label}
For attribution, please cite this work as:
:::

::: {#ref-csefalvay2023 .csl-entry .quarto-appendix-citeas}
Chris von Csefalvay. 2023. "The Lyre of Hephaestus."
<https://doi.org/10.59350/pd0q6-r5n60>.
:::

</div>
:::
:::